SPORTS MATH

An Introductory Course in the Mathematics

of Sports Science and Sports Analytics
 TEXTBOOKS in MATHEMATICS
Series Editors: Al Boggess and Ken Rosen

PUBLISHED TITLES
ABSTRACT ALGEBRA: AN INTERACTIVE APPROACH, SECOND EDITION
William Paulsen
ABSTRACT ALGEBRA: AN INQUIRY-BASED APPROACH
Jonathan K. Hodge, Steven Schlicker, and Ted Sundstrom
ADVANCED LINEAR ALGEBRA
Hugo Woerdeman
APPLIED ABSTRACT ALGEBRA WITH MAPLE™ AND MATLAB®, THIRD EDITION
Richard Klima, Neil Sigmon, and Ernest Stitzinger
APPLIED DIFFERENTIAL EQUATIONS: THE PRIMARY COURSE
Vladimir Dobrushkin
COMPUTATIONAL MATHEMATICS: MODELS, METHODS, AND ANALYSIS WITH MATLAB® AND MPI,
SECOND EDITION
Robert E. White
DIFFERENTIAL EQUATIONS: THEORY, TECHNIQUE, AND PRACTICE, SECOND EDITION
Steven G. Krantz
DIFFERENTIAL EQUATIONS: THEORY, TECHNIQUE, AND PRACTICE WITH BOUNDARY VALUE PROBLEMS
Steven G. Krantz
DIFFERENTIAL EQUATIONS WITH MATLAB®: EXPLORATION, APPLICATIONS, AND THEORY
Mark A. McKibben and Micah D. Webster
ELEMENTARY NUMBER THEORY
James S. Kraft and Lawrence C. Washington
EXPLORING CALCULUS: LABS AND PROJECTS WITH MATHEMATICA®
Crista Arangala and Karen A. Yokley
EXPLORING LINEAR ALGEBRA: LABS AND PROJECTS WITH MATHEMATICA®
Crista Arangala
GRAPHS & DIGRAPHS, SIXTH EDITION
Gary Chartrand, Linda Lesniak, and Ping Zhang
PUBLISHED TITLES CONTINUED

INTRODUCTION TO ABSTRACT ALGEBRA, SECOND EDITION

Jonathan D. H. Smith
INTRODUCTION TO MATHEMATICAL PROOFS: A TRANSITION TO ADVANCED MATHEMATICS, SECOND EDITION
Charles E. Roberts, Jr.
INTRODUCTION TO NUMBER THEORY, SECOND EDITION
Marty Erickson, Anthony Vazzana, and David Garth
LINEAR ALGEBRA, GEOMETRY AND TRANSFORMATION
Bruce Solomon
MATHEMATICAL MODELLING WITH CASE STUDIES: USING MAPLE™ AND MATLAB®, THIRD EDITION
B. Barnes and G. R. Fulford
MATHEMATICS IN GAMES, SPORTS, AND GAMBLING–THE GAMES PEOPLE PLAY, SECOND EDITION
Ronald J. Gould
THE MATHEMATICS OF GAMES: AN INTRODUCTION TO PROBABILITY
David G. Taylor
A MATLAB® COMPANION TO COMPLEX VARIABLES
A. David Wunsch
MEASURE THEORY AND FINE PROPERTIES OF FUNCTIONS, REVISED EDITION
Lawrence C. Evans and Ronald F. Gariepy
NUMERICAL ANALYSIS FOR ENGINEERS: METHODS AND APPLICATIONS, SECOND EDITION
Bilal Ayyub and Richard H. McCuen
ORDINARY DIFFERENTIAL EQUATIONS: AN INTRODUCTION TO THE FUNDAMENTALS
Kenneth B. Howell
RISK ANALYSIS IN ENGINEERING AND ECONOMICS, SECOND EDITION
Bilal M. Ayyub
SPORTS MATH: AN INTRODUCTORY COURSE IN THE MATHEMATICS OF SPORTS SCIENCE AND
SPORTS ANALYTICS
Roland B. Minton
TRANSFORMATIONAL PLANE GEOMETRY
Ronald N. Umble and Zhigang Han
 TEXTBOOKS in MATHEMATICS

SPORTS MATH
An Introductory Course in the Mathematics
of Sports Science and Sports Analytics

Roland B. Minton
Roanoke College
Salem, Virginia, USA
CRC Press
Taylor & Francis Group
6000 Broken Sound Parkway NW, Suite 300
Boca Raton, FL 33487-2742
© 2017 by Taylor & Francis Group, LLC
CRC Press is an imprint of Taylor & Francis Group, an Informa business

No claim to original U.S. Government works

Printed on acid-free paper

Version Date: 20160816

International Standard Book Number-13: 978-1-4987-0626-1 (Hardback)

This book contains information obtained from authentic and highly regarded sources. Reasonable
efforts have been made to publish reliable data and information, but the author and publisher cannot
assume responsibility for the validity of all materials or the consequences of their use. The authors and
publishers have attempted to trace the copyright holders of all material reproduced in this publication
and apologize to copyright holders if permission to publish in this form has not been obtained. If any
copyright material has not been acknowledged please write and let us know so we may rectify in any
future reprint.

Except as permitted under U.S. Copyright Law, no part of this book may be reprinted, reproduced,
transmitted, or utilized in any form by any electronic, mechanical, or other means, now known or
hereafter invented, including photocopying, microfilming, and recording, or in any information stor-
age or retrieval system, without written permission from the publishers.

For permission to photocopy or use material electronically from this work, please access www.copy-
right.com (http://www.copyright.com/) or contact the Copyright Clearance Center, Inc. (CCC), 222
Rosewood Drive, Danvers, MA 01923, 978-750-8400. CCC is a not-for-profit organization that pro-
vides licenses and registration for a variety of users. For organizations that have been granted a photo-
copy license by the CCC, a separate system of payment has been arranged.

Trademark Notice: Product or corporate names may be trademarks or registered trademarks, and are
used only for identification and explanation without intent to infringe.
Visit the Taylor & Francis Web site at
http://www.taylorandfrancis.com
and the CRC Press Web site at
http://www.crcpress.com
Contents

Preface xiii

List of Figures xvii

List of Tables xxi

1 Projectile Motion 1

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
Figuring with Newton . . . . . . . . . . . . . . . . . . . . . . . . . 1
Hangin’ with MJ: 1-D Motion . . . . . . . . . . . . . . . . . . . . 3
Raining 3’s with Steph: 2-D Motion . . . . . . . . . . . . . . . . . 4
K’s with Kershaw: Terminal Velocity and Drag Forces . . . . . . . 6
Calculus Box: Solving for Velocity . . . . . . . . . . . . . . . 7
Bending with Bubba: Magnus Force . . . . . . . . . . . . . . . . . 8
Smiling with Dimples . . . . . . . . . . . . . . . . . . . . . . . . . 11
Calculus Box: A General Model of a Ball in Flight . . . . . . 12
The Effects of Drag and Lift . . . . . . . . . . . . . . . . . . . . . 13
Knuckling Down . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
Calculus Box: Lateral Position of a Knuckleball . . . . . . . . 17
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
Further Reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

2 Rotational Motion 25

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
Going in Circles . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
Torquing Off Newton . . . . . . . . . . . . . . . . . . . . . . . . . 28
All About MOI . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
Size Is Important . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
Calculus Box: Calculating MOI . . . . . . . . . . . . . . . . . 30
Equipment Design . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
Supercats and Tamedogs . . . . . . . . . . . . . . . . . . . . . . . 32
Keeping the Momentum . . . . . . . . . . . . . . . . . . . . . . . 33
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
Further Reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38

vii
viii Contents

3 Sports Illusions 39

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
You Can’t Keep Your Eye on the Ball . . . . . . . . . . . . . . . . 39
You Can’t Touch This . . . . . . . . . . . . . . . . . . . . . . . . . 41
You Can’t Teach Size . . . . . . . . . . . . . . . . . . . . . . . . . 42
You Can’t Afford the Yardage . . . . . . . . . . . . . . . . . . . . 43
You Can’t Bend That Way . . . . . . . . . . . . . . . . . . . . . . 45
You Can’t Make That Call! . . . . . . . . . . . . . . . . . . . . . . 46
You Can’t Clear That Bar . . . . . . . . . . . . . . . . . . . . . . 49
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
Further Reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54

4 Collisions 55

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
Linear Momentum . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
Impulse and Force . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
Calculus Box: Integration . . . . . . . . . . . . . . . . . . . . . . . 59
Giving to Receive . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
Tendons and Tennis . . . . . . . . . . . . . . . . . . . . . . . . . . 61
Coefficient of Restitution . . . . . . . . . . . . . . . . . . . . . . . 62
Incoming and Outgoing . . . . . . . . . . . . . . . . . . . . . . . . 63
Derivative Works . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
The Way the Ball Bounces . . . . . . . . . . . . . . . . . . . . . . 66
Freeze Frame . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
Further Reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71

5 Ratings Systems 73

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
Right versus Best . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
Ratings versus Rankings . . . . . . . . . . . . . . . . . . . . . . . 74
The Massey System . . . . . . . . . . . . . . . . . . . . . . . . . . 75
Connected Schedules . . . . . . . . . . . . . . . . . . . . . . . 77
Massey Win Ratings . . . . . . . . . . . . . . . . . . . . . . . 78
Offense and Defense . . . . . . . . . . . . . . . . . . . . . . . 79
Least Squares Equivalence . . . . . . . . . . . . . . . . . . . . 79
Wins versus Points . . . . . . . . . . . . . . . . . . . . . . . . . . 80
The Colley System . . . . . . . . . . . . . . . . . . . . . . . . . . 81
A Flaky Scaling Problem . . . . . . . . . . . . . . . . . . . . 82
The Elo System . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
Strength of Schedule . . . . . . . . . . . . . . . . . . . . . . . . . 85
Computing Probabilities . . . . . . . . . . . . . . . . . . . . . . . 85
 Contents ix

Weighty Issues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
Calculus Box: A Recipe for Reduction of Matrices . . . . . . . . . 87
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
Further Reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94

6 Voting Systems 95

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
How They Vote . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
Condorcet’s Intransitive Attitude . . . . . . . . . . . . . . . . . . 97
Preference Lists, Voting Systems, and Chaos . . . . . . . . . . . . 99
Fairness and the Arrow of Impossibility . . . . . . . . . . . . . . . 101
Positional Voting Systems . . . . . . . . . . . . . . . . . . . . . . 102
A Return to Sports Voting . . . . . . . . . . . . . . . . . . . . . . 103
Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
Range Voting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
PageRank and MVPassing . . . . . . . . . . . . . . . . . . . . . . 108
Seeding of Tournaments . . . . . . . . . . . . . . . . . . . . . . . . 110
Probability Box: Put Some Error Bars on Those Things . . . . . . 112
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
Further Reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119

7 Saber- and Other Metrics 121

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
The Pythagorean Cult . . . . . . . . . . . . . . . . . . . . . . . . 122
When Good Statistics Go Bad . . . . . . . . . . . . . . . . . . . . 125
Rates versus Numbers . . . . . . . . . . . . . . . . . . . . . . . . . 127
Persistence and Reliability . . . . . . . . . . . . . . . . . . . . . . 128
On the Defensive . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
Plus and Minus . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
Park Factors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
Four Factors, Fenwick, and Football . . . . . . . . . . . . . . . . . 134
Evaluation and Prediction . . . . . . . . . . . . . . . . . . . . . . 137
Regression to the Mean . . . . . . . . . . . . . . . . . . . . . . . . 138
Linear Weights: A Prelude to WAR . . . . . . . . . . . . . . . . . 139
Calculus Box: Linear Regression . . . . . . . . . . . . . . . . 141
Roger Maris and the Hall of Fame . . . . . . . . . . . . . . . . . . 141
Now Trending . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
Further Reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
x Contents

8 Randomness in Sports 151

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
Summing Up the Basics . . . . . . . . . . . . . . . . . . . . . . . . 152
Prediction is Difficult . . . . . . . . . . . . . . . . . . . . . . . . . 154
A Slump or a Disaster . . . . . . . . . . . . . . . . . . . . . . . . . 155
Calculus Box: Probability . . . . . . . . . . . . . . . . . . . . . . . 156
May the Best Team Win . . . . . . . . . . . . . . . . . . . . . . . 157
Measuring Parity: Gini in a Bottle . . . . . . . . . . . . . . . . . . 158
Measuring Parity: Luck versus Skill . . . . . . . . . . . . . . . . . 161
The Paradox of Skill . . . . . . . . . . . . . . . . . . . . . . . . . . 162
Measuring Parity: Entropy . . . . . . . . . . . . . . . . . . . . . . 163
Declaration of Independence . . . . . . . . . . . . . . . . . . . . . 164
Conditional Probability . . . . . . . . . . . . . . . . . . . . . . . . 165
The Hot Hands . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
Not So Fast, My Friend . . . . . . . . . . . . . . . . . . . . . . . . 167
Runs Tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
Joltin Joe and The Streak . . . . . . . . . . . . . . . . . . . . . . 169
Not Following the Rules . . . . . . . . . . . . . . . . . . . . . . . . 170
BABIP and DIPS . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
Random Thoughts . . . . . . . . . . . . . . . . . . . . . . . . . . . 172
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
Further Reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . 178

9 Sports Strategies 179

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
Don’t Punt, John! . . . . . . . . . . . . . . . . . . . . . . . . . . . 180
Bill Belichick’s Gambles . . . . . . . . . . . . . . . . . . . . . . . . 181
The Value of a Play . . . . . . . . . . . . . . . . . . . . . . . . . . 183
Markov Chain Models . . . . . . . . . . . . . . . . . . . . . . . . . 184
The Expected Runs Matrix . . . . . . . . . . . . . . . . . . . . . . 187
Win Probability and Leverage . . . . . . . . . . . . . . . . . . . . 187
Game Control and the Story Stat . . . . . . . . . . . . . . . . . . 188
Game Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189
Upsetting the Game Theory . . . . . . . . . . . . . . . . . . . . . 193
Getting and Giving Two . . . . . . . . . . . . . . . . . . . . . . . 194
The Physical Challenge . . . . . . . . . . . . . . . . . . . . . . . . 195
Personnel Decisions: Aging . . . . . . . . . . . . . . . . . . . . . . 196
Personnel Decisions: Transfer Fees and Stars . . . . . . . . . . . . 197
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
Further Reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202
 Contents xi

10 Big Data and Beyond 203

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
Big Data Is Watching You . . . . . . . . . . . . . . . . . . . . . . 204
A Theory of Everything . . . . . . . . . . . . . . . . . . . . . . . . 205
Catch Me If You Can . . . . . . . . . . . . . . . . . . . . . . . . . 207
Getting Framed . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208
Anonymous Field Goal Kicking . . . . . . . . . . . . . . . . . . . 209
On the Rebound . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211
Breaching the Convex Hull . . . . . . . . . . . . . . . . . . . . . . 213
Calculus Box: A Goal-Scoring Model . . . . . . . . . . . . . . . . 215
Showing Hot and Cold . . . . . . . . . . . . . . . . . . . . . . . . 217
RIP to the RPI . . . . . . . . . . . . . . . . . . . . . . . . . . . . 218
Blackbox Analytics . . . . . . . . . . . . . . . . . . . . . . . . . . 220
PeeWee Analytics . . . . . . . . . . . . . . . . . . . . . . . . . . . 222
Wearable Tech . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223
Further Reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . 226

Answers and Selected Solutions 227

Index 253
Preface

This is a textbook for a course that does not exist. Like the myths that are
busted in Chapter 3, that statement has some truth to it. As Obi Wan Kenobi
would say, it is true “from a certain point of view.” There are a number of
existing courses with Sports Science and Sports Analytics in the titles, created
by intrepid professors venturing into the unknown. The topics and emphases
of such courses vary dramatically, so that there is no consensus on what a
course in Sports Science and/or Sports Analytics should be.
There are conferences on sports analytics. The MIT/ESPN Sloan Sports
Analytics Conferences are graced with outstanding speakers, and the demand
for tickets grows exponentially. The topics at a conference can range from
Moneyball to marketing strategies, management strategies or technological
breakthroughs. Regional conferences such as the Carolina Sports Analytics
Meeting provide support for the increasingly large number of faculty and
students doing research in sports-related areas.
To use a golfing analogy, writing a book like this is like hitting a drive
at a driving range; there are many directions you can go without going out
of bounds. At the driving range, I pick out a small target to focus on, and
that is what I have done here. I have chosen a sample of topics that I know
something about and that I find very interesting. Ideally, users of this book
will have enough to choose from to suit whichever version of a sports course
is being run.
The course that I have taught at Roanoke College since 1988 is a mix
of physics, physiology, mathematics, and statistics. The order (and level of
emphasis) of the topics has changed over the years; this book reflects the
current status of my course. It is, admittedly, an eclectic mix of topics (at the
driving range, I may aim at one target, but I do tend to spray balls all over the
range). I hope to provide ideas and resources to help students launch projects.
An important part of my course is the term project, and I have almost always
been pleasantly surprised at the quality of work done in a short period of time.
I suspect that the high quality of work is due to the students’ high level
of motivation; not from any talents of mine, but because many students (of
both genders) find it exciting to think about sports and to complete a research
agenda. Sports problems are easy to create and state, even for students who
do not live sports 24/7. Sports are part of their culture and knowledge base,
and the opportunity to be an expert on some area of sports is invigorating.

xiii
xiv Preface

This should be the primary reason for the growth of sports courses: the topic
provides intrinsic motivation for students to do their best work.
This, as I said, is a textbook. That fact alters the literary qualities of
the writing. My intention is for it to be easy and enjoyable to read, but
examples and exercises necessarily interrupt the normal flow of text. As well,
the exercises guide students to some very interesting results, so that some of
the best discoveries about sports may be hiding in the exercises. I encourage
you to look for fun facts in the exercises.
The choice of mathematical level is problematic for a book like this. Some
of my favorite results require calculus or even differential equations for a full
explanation, but I do not want to narrow the audience to the mathematically
advanced. I have split the difference on calculus. I am not assuming that you
know calculus, but I will show you some of the things that calculus can do
for you. Those of you who have taken calculus can read the “calculus box”
sections in the text and work the exercises labeled as calculus exercises. If you
have not taken calculus, simply navigate around those well-marked areas of
the book.
The extent to which a background in probability and statistics is required
is more difficult to say. Sports analytics relies heavily on sound statistical rea-
soning. Statistical “common sense” is assumed throughout, but the details of
tests and calculations are all provided. Similarly, a familiarity with the ideas
of computing is assumed, but no programming is required. The reader’s expe-
rience will be greatly enhanced by frequent use of the internet, spreadsheets,
and calculations.
I should admit that I like to read books; I enjoy holding physical books.
On the other hand, I now buy most of my books and music in digital format.
And I am slowly allowing myself to stream a movie or music online and let
it slip away without claiming possession. The point of this ramble is that
while I recognize that the future of sports research is digital with remote
access, this book has a fairly standard format. There will be a website at
www.roanoke.edu/mcsp/minton/SportsMath.html (I know, I’m showing my
age by posting a url that will change. A search for “Minton Sports Math”
should do it, but you don’t need me to tell you that). I’ll post links, references,
notes, and anything else that comes to mind that could be useful and does
not fit the classic book mold. Ideally, part of the site will even be wiki-like.
In the last thirty years, data collection has progressed from repeated view-
ings of grainy videos to nearly continuous data streaming from sensors at-
tached to every part of an athlete, from Bill James painstakingly copying box
score numbers from The Sporting News to a one-minute online search that
lists the top fifty hitting streaks in MLB history. My hope is that this book
opens up some of the astounding possibilities of sports research, while helping
you learn more about the games you enjoy.
This book would not exist without the encouragement of my editor, Bob
Ross, who over the years has furthered my career in multiple ways. Thanks,
Bob! My Roanoke College family has provided support in several ways. Dave
 Preface xv

Taylor has listened to countless musings and rants on all aspects of the book,
and provided good counsel at all times. His assistance with the joys of TeX
is invaluable. Adam Childers provided much-needed statistical backing, plus
hours of enjoyable sports talk. Thanks to Karin Saoub and Chris Lee for their
assistance. The athletic department, especially Ryan Pflugrad, Matt McGuire,
Page Moir, Scott Allison, and Chris Kilcoyne are great to work with. An
important chunk of the time to do this enjoyable work was provided by the
M. Paul Capp and Constance Whitehead Endowed Chair, for which I am very
grateful. Paul is a great supporter of education, especially in mathematics and
physics. Thanks to Dean Richard Smith for his support; it is very cool to get to
cite my Dean’s publication in this book! Finally, to my wife Jan and children
Kelly and Greg, who deal with me in writing mode, which is even grumpier
than usual: thanks for being who you are, and for your love.

Dr. Roland Minton

MCSP Department
Roanoke College
Salem, VA 24153
minton@roanoke.edu
List of Figures

1.1 Height vs Time Graph for Jump . . . . . . . . . . . . . . . . 3

1.2 Components of Initial Velocity . . . . . . . . . . . . . . . . . 4
1.3 Force on Spinning Ball . . . . . . . . . . . . . . . . . . . . . 9
1.4 Spin Vector . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.5 RightHand Rule . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.6 Magnus Force Up . . . . . . . . . . . . . . . . . . . . . . . . 10
1.7 Magnus Force Up and Back . . . . . . . . . . . . . . . . . . 10
1.8 Magnus Force Up and Forward . . . . . . . . . . . . . . . . 10
1.9 Effect of Drag and Magnus Forces . . . . . . . . . . . . . . . 14
1.10 Effect of Spin on Pitches . . . . . . . . . . . . . . . . . . . . 15
1.11 Knuckleball with No Spin . . . . . . . . . . . . . . . . . . . 16
1.12 Knuckleball with 1/4 Rotation . . . . . . . . . . . . . . . . . 16
1.13 Knuckleball with 1 Rotation . . . . . . . . . . . . . . . . . . 16

2.1 Polar Coordinates . . . . . . . . . . . . . . . . . . . . . . . . 26

3.1 Keep Your Eye on the Ball . . . . . . . . . . . . . . . . . . . 40

3.2 Field Goal Angles . . . . . . . . . . . . . . . . . . . . . . . . 43
3.3 Golf Club Bend . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.4 The Offside Call . . . . . . . . . . . . . . . . . . . . . . . . . 46
3.5 Optical Errors on Offside Call . . . . . . . . . . . . . . . . . 47
3.6 Close Play at First Base . . . . . . . . . . . . . . . . . . . . 48
3.7 Olympic Winning Heights . . . . . . . . . . . . . . . . . . . 50
3.8 Clearing the Bar . . . . . . . . . . . . . . . . . . . . . . . . 50

4.1 Area for Nonconstant Force . . . . . . . . . . . . . . . . . . 58

4.2 Area for Nonlinear Force . . . . . . . . . . . . . . . . . . . . 58
4.3 Force versus Displacement . . . . . . . . . . . . . . . . . . . 61
4.4 Force versus Displacement . . . . . . . . . . . . . . . . . . . 61
4.5 Speeds Before and After . . . . . . . . . . . . . . . . . . . . 63

5.1 Elo Performace Graph . . . . . . . . . . . . . . . . . . . . . 84

5.2 Histogram of Predicted Minus Actual Scores, 2014 . . . . . 86

6.1 The Author with Condorcet Statue in Paris . . . . . . . . . 98

6.2 Borda Count, Approval, Plurality Simulation . . . . . . . . 105

xvii
xviii List of Figures

6.3 Two Controversial Candidates Simulation . . . . . . . . . . 106

6.4 Range vs Borda . . . . . . . . . . . . . . . . . . . . . . . . . 107
6.5 First Round Wins by Seed . . . . . . . . . . . . . . . . . . . 111
6.6 Predicted Wins by Seed, Rounds 1-4 . . . . . . . . . . . . . 111
6.7 Percentage in Final Four by Seed . . . . . . . . . . . . . . . 112

7.1 Bill James . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121

7.2 The Best Exponent, MLB 2014 . . . . . . . . . . . . . . . . 123
7.3 BA vs OBP, 2004 . . . . . . . . . . . . . . . . . . . . . . . . 125
7.4 Runs vs OPS, 2014 . . . . . . . . . . . . . . . . . . . . . . . 126
7.5 BA vs OBP, 2014 . . . . . . . . . . . . . . . . . . . . . . . . 127
7.6 Average Score versus Length of Hole . . . . . . . . . . . . . 139
7.7 MLB Strikeouts, 2000-2014 . . . . . . . . . . . . . . . . . . . 142
7.8 NFL Passing Yards, 2000-2014 . . . . . . . . . . . . . . . . . 143
7.9 NBA Three-Point Attempts, 2000-2014 . . . . . . . . . . . . 143
7.10 EPL Goals, 2000-2014 . . . . . . . . . . . . . . . . . . . . . 144

8.1 A Behind the Back Shot . . . . . . . . . . . . . . . . . . . . 151

8.2 Points Scored by LA Clippers, 2014-15 . . . . . . . . . . . . 153
8.3 Points Scored by LA Clippers, 2014-15 and Normal Curve . 153
8.4 Points per Game by Players, 2014-15 . . . . . . . . . . . . . 153
8.5 Lorenz Curve for Perfect Parity . . . . . . . . . . . . . . . . 159
8.6 Lorenz Curve for Unequal League . . . . . . . . . . . . . . . 159
8.7 Lorenz Curve for Two Leagues . . . . . . . . . . . . . . . . . 159
8.8 Area Defining the Gini Index . . . . . . . . . . . . . . . . . 160
8.9 Lorenz Curve for Example 8.4 . . . . . . . . . . . . . . . . . 160
8.10 Skill Curve and Maximum . . . . . . . . . . . . . . . . . . . 162
8.11 Skill Curve Approaching Maximum . . . . . . . . . . . . . . 162

9.1 Kevin Kelley . . . . . . . . . . . . . . . . . . . . . . . . . . . 179

9.2 Win Probabilities for Two Baseball Games . . . . . . . . . . 189
9.3 Home Runs by Age, 2014 AL . . . . . . . . . . . . . . . . . 196
9.4 Home Run Rate by Age, 2014 AL . . . . . . . . . . . . . . . 196
9.5 Best 3 WARs versus Wins . . . . . . . . . . . . . . . . . . . 198
9.6 Worst 3 WARs versus Wins . . . . . . . . . . . . . . . . . . 198

10.1 Jump Throw to First . . . . . . . . . . . . . . . . . . . . . . 203

10.2 Rebounders in Position . . . . . . . . . . . . . . . . . . . . . 211
10.3 Rebounders with Dividing Lines . . . . . . . . . . . . . . . . 212
10.4 Voronoi Diagram for Rebounders . . . . . . . . . . . . . . . 212
10.5 Two Defensive Alignments . . . . . . . . . . . . . . . . . . . 214
10.6 Convex Hulls of Two Defensive Alignments . . . . . . . . . 214
10.7 Goals in 2014-15 EPL Games . . . . . . . . . . . . . . . . . 216
10.8 Home/Away Goals in 2014-15 EPL Games . . . . . . . . . . 216
10.9 Heat Maps for Batting Averages . . . . . . . . . . . . . . . . 217
10.10 Locations of Pitches: Bad Graphic . . . . . . . . . . . . . . 218
 List of Figures xix

10.11 Locations of Pitches: Good Graphic . . . . . . . . . . . . . . 218

10.12 Basic Neural Network Diagram . . . . . . . . . . . . . . . . 220
10.13 Golf Drive Data . . . . . . . . . . . . . . . . . . . . . . . . . 222
List of Tables

1.1 Formulas for Constant Acceleration . . . . . . . . . . . . . . 3

1.2 Heights and Times for Jump . . . . . . . . . . . . . . . . . . 4

2.1 Formulas for Constant Acceleration . . . . . . . . . . . . . . 26

2.2 Translational to Rotational . . . . . . . . . . . . . . . . . . 26

5.1 Four-Team League Results . . . . . . . . . . . . . . . . . . . 76

5.2 Four-Team League Schedule . . . . . . . . . . . . . . . . . . 76
5.3 Top 5 NCAA Points . . . . . . . . . . . . . . . . . . . . . . 78
5.4 Top 5 NCAA Wins . . . . . . . . . . . . . . . . . . . . . . . 78
5.5 Top 5 NCAA Offense . . . . . . . . . . . . . . . . . . . . . . 79
5.6 Top 5 NCAA Defense . . . . . . . . . . . . . . . . . . . . . . 79
5.7 Colley Ratings . . . . . . . . . . . . . . . . . . . . . . . . . . 82
5.8 Colley Double Ratings . . . . . . . . . . . . . . . . . . . . . 82

8.1 Gini Indices for Leagues . . . . . . . . . . . . . . . . . . . . 161

8.2 Entropy Values for Leagues . . . . . . . . . . . . . . . . . . 164
8.3 Simulated and Actual Hitting Streaks . . . . . . . . . . . . . 170

10.1 Large vs Small Conference RPI . . . . . . . . . . . . . . . . 219

xxi
Chapter 1
Projectile Motion

Introduction
Basketball star Stephen Curry launches a
3-point shot. As the ball traces its high arc
toward the basket, fans rise to their feet in
anticipation. Will it go in? Is it a little short?
Similar tension accompanies a Jordan Spieth
tee shot, an Andy Murray passing shot, a
long football pass by Peyton Manning or Li-
onel Messi, or a long fly ball by Mike Trout.
We will analyze the flights of balls in this
chapter as we explore the area of physics
known as mechanics.
Along the way, we will answer such ques-
tions as: How does Blake Griffin hang in the
air when dunking? What is the optimal angle
to shoot a free throw? Why do golf balls have
dimples? Does a knuckleball really dance?
The answers are to be found in the funda-
mentals of physics.

Figuring with Newton

Sir Isaac Newton (1643-1727) constructed a framework for the analysis of
objects in motion. The second of his three Laws of Motion is the launching
point for most of our investigations in this chapter. The shorthand version of
Newton’s Second Law is
F = ma
where F is the sum of all forces acting on an object, m is the object’s mass,
and a is the acceleration of the object. One of the most remarkable aspects of

1
2 Sports Math

Newton’s Second Law is that it can also be written as F = ma, where F and
a appear in bold to indicate that they are multidimensional vector quantities.
We will return to this form of the equation when we look at motion in two
and three dimensions. The mass m is a scalar (real number) that is related
to weight: for earthbound sports, weight is approximately equal to mass times
the gravitational constant g.
To keep it simple, let’s start with one-dimensional motion; vertical motion,
to be precise. In this case, the object’s position can be tracked by its height h
above some reference point (e.g., the ground). We define velocity as the rate
of change of position with respect to time. At a constant speed, this means
∆h
that velocity equals change in height divided by change in time: v = . This
∆t
gets complicated when velocity is not constant. In general,
∆h
Average velocity =
∆t
and, for small time intervals, (instantaneous) velocity is approximately equal
∆h
to average velocity: v ≈ . With calculus, we can simply say that velocity
∆t
is the derivative of height. Either way, note that v can be negative (if height
is decreasing) or positive (if height is increasing). The acceleration a of the
∆v
object is, in turn, the rate of change of velocity. Then a ≈ and acceleration
∆t
is the derivative of velocity.

Example 1.1 Suppose a ball falls from a height of 50 meters. If gravity is

the only force on the ball, find the velocity of the ball after t = 1 second and
t = 1.5 seconds.
Solution. For most sports situations, we can assume that the acceleration due
to gravity is a constant −g with g ≈ 9.8 m/s2 or g ≈ 32 ft/s2 . An acceleration
of 9.8 m/s2 in the negative direction means that in every second the velocity
decreases by 9.8 m/s. Assuming that the ball starts with velocity 0, then at
t = 1 second the velocity has decreased to −9.8 m/s. In the next half-second,
the velocity decreases by 0.5(9.8) m/s = 4.9 m/s. At time t = 1.5 s the velocity
has decreased to (−9.8 − 4.9) m/s = −14.7 m/s. The ideas from this basic
example will be used again for the more complicated situation of Figure 1.9.

Speed is defined as the absolute value of velocity. In Example 1.1 above, at

time t = 1 the ball’s velocity is −9.8 m/s but its speed is 9.8 m/s (downward).
Notice that Example 1.1 did not ask for heights. Because the ball’s velocity
is changing, the calculation of position from velocity requires more than mul-
tiplying velocity by time. Fortunately, calculus gives us some simple formulas
to use, shown below in Table 1.1.
In Example 1.1, we have c = −g, v0 = 0, and p0 = 50, so the height at
time t is −4.9t2 + 50 m. At t = 1, the ball is at height −4.9 + 50 m = 45.1 m,
while at t = 1.5 the ball is at height −4.9(1.5)2 + 50 m = 38.975 m.
 Projectile Motion 3
TABLE 1.1: Formulas for Constant Acceleration
acceleration a=c
velocity v = ct + v0
position p = 12 ct2 + v0 t + p0

Hangin’ with MJ: 1-D Motion

Using the equations in Table 1.1, we can discover an interesting fact about
vertical motion. We start with a straightforward calculation.

Example 1.2 A man jumps from the ground with an initial velocity of 16
ft/s, under the force of gravity. (a) How long does he stay in the air? (b) How
high does he go?
Solution. We use Table 1.1 with c = −32, v0 = 16, and p0 = 0. (Note
that gravity pulls in the negative direction, while the jump is in the positive
direction.) Then velocity is v = −32t + 16 ft/s and position is h = −16t2 + 16t
ft. Now, let’s decipher the questions being asked. (a) What does “in the air”
mean? He is in the air from launch time (height 0) to landing time (height
0). Both times occur at height 0, when h = −16t2 + 16t = 0. So, solve this
equation! If −16t(t − 1) = 0, then t = 0 or t = 1. He launches at t = 0 and
lands at t = 1, hence is in the air for 1 second. (b) At the top of a jump,
velocity is 0: no longer going up, not yet coming down. This occurs when
v = −32t + 16 = 0 or t = 21 . Now that we know when he reaches his peak,
we can determine his height using the position function. The height at time

2
t = 12 is h = −16 12 + 16 12 = −4 + 8 = 4 feet.

The solution of Example 1.2 follows a pattern that you should use in most
such problems. First, get the equations of motion by filling in the constants in
Table 1.1. Then, solve one of the equations for time t based on the situation
(e.g., how long the object is in the air, or when it reaches its peak). Finally,
substitute this time value into another equation to find the quantity of interest.
The 48-inch jump of Example 1.2 is in leg-
endary leaper status, up there with Michael
Jordan and Blake Griffin. But, why do these
prodigious leapers seem to hang in the air?
One reason is that all objects hang in the air.
The graph of height versus time in Figure 1.1
and Table 1.2 below show the height for the
jumper in Example 1.2 at equal quarter marks
in time. FIGURE 1.1: Jump
Notice that from time t = 1/4 to t = 3/4 (which is half of the time of the
jump) the height is 3 feet or above (with a peak height of 4 feet). That is,
4 Sports Math
TABLE 1.2: Heights and Times for Jump
Time (s) Height (ft)
0 0
1/4 3
1/2 4
3/4 3
1 0

half the time is spent in the top one-quarter of the jump! The speed
is smallest at the top of the flight, so the object “hangs” at the top.
A second reason that great athletes can appear to defy gravity has to do
with center of mass. The center of mass is where the sums of mass-times-
distance quantities balance. For a standing human being, it is not far from the
geometric center of the body. Newton’s equations track the center of mass of
the object in flight. Figure 1.1 does not show a body in flight, but the path of
a single point. That point is the center of mass of the person. (Which means
that a “height” of 0 does not actually mark the location of the ground; it
marks the location of the center of mass of the object at launch time.) While
the dunker’s center of mass is tracking the nice parabola shown, he is free to
pull up his legs, bob his head, and extend an arm in entertaining ways that
may cause an individual body part such as the head to remain at the same
height for a noticeable amount of time.

Raining 3’s with Steph: 2-D Motion

Let’s return to Stephen Curry’s 3-point shot. We can analyze its flight with
Newton’s Second Law, but the fact that the ball now moves both horizontally
and vertically complicates the calculations.
From nba.com/Stats, we can get an idea of the location of Curry’s shot. In
2014-15, only 79 of Curry’s 618 3-pointers were from the corners. (Remarkably,
he made well over 40% of his shots from every 3-point zone and 62% from the
left corner, plus an outrageous 91% from the left corner during the playoffs.)
Most of his shots were from beyond the arc that is 23.75 feet from the basket.
Let’s say his shot is from 25 feet away. Align the x-axis horizontally from
Curry to the basket, and the y-axis vertically.
We will assume that Curry’s impeccable form
keeps the ball from curving left or right. Newton’s
Second Law is the vector equation F = ma where
the vectors F and a have two components. That is,
the acceleration has a horizontal component ax and
a vertical component ay . Assuming that gravity

FIGURE 1.2: Velocity

 Projectile Motion 5

is the only force, then ay = −g as before, and ax = 0 (no forces acting

horizontally). This allows us to separate the x- and y-equations. To use Table
1.1, we need the initial velocities and initial positions. We assume that p0x = 0
ft for convenience and p0y = 7 ft (assuming the ball is released from a height
of 7 feet). If the ball is launched with speed 30 ft/s at an angle of 50 degrees,
then v0x and v0y are obtained from the triangle in Figure 1.2.
Using basic trigonometry, we get initial velocities v0x = 30 cos(50◦ ) ft/s
and v0y = 30 sin(50◦ ) ft/s, or v0x ≈ 19.28 ft/s and v0y ≈ 22.98 ft/s. Pulling
this all together, we have x ≈ 19.28t and y ≈ −16t2 + 22.98t + 7.

Example 1.3 Is this shot good or not?

Solution. In this case, a perfect shot would pass through x = 25 and y = 10
(the height of the basket). We will solve for t in one equation and plug into
the other equation, but that can be done in two ways. For reasons you will
see, it is more convenient to start with the y-equation. We want y = 10 and so
solve −16t2 + 22.98t + 7 = 10 for t. There are two solutions, one representing
the ball rising up through the height y = 10 and the other representing the
ball dropping through the height y = 10; the second solution is clearly the
one of interest. We get t ≈ 1.29 s. If the shot is perfect, then at this t-value
we get x = 25 (be sure this makes sense to you!). Instead, our equation gives
x ≈ 24.90 feet. Not perfect, but is this close enough? The center of the basket
is at x = 25, so x = 24.90 represents 0.1 foot or 1.2 inches from the center.
The basket has diameter 18 inches and the ball has diameter 9.5 inches, so
the ball can move a little over 4 inches from the center and still be inside the
basket. (This assumes that the shot is exactly on line.) Count the three!

The work in Example 1.3 does not fully prove that the shot is good. Can
you think of what is missing?
Even if the center of the ball (theoretically) passes inside the basket, in
real life if the trajectory of the ball is too flat some portion of the ball will
hit the rim. You will show in exercise 1.41 that the ball in Example 1.3 enters
the basket at an angle of about 43 degrees, more than steep enough to safely
pass through the basket.
We can now develop a method to determine the best angle at which to
shoot a free throw. An important part of our interpretation of the numbers
in Example 1.3 is the margin of error inherent in playing with a ball that
is smaller than the basket. You could imagine decreasing the initial speed
from 30 ft/s until the shot is no longer good; call this speed s1 . Then find the
maximum speed s2 for which the shot is good. For the angle 50 degrees, s2 −s1
is the margin of error in speed. The bigger the margin of error, the better,
since the shooter does not have to be as precise with the launch speed. Peter
Brancazio has done this study and found that a free throw angle of about 49
degrees gives the largest margin of error. We will explore an interesting aspect
of this angle in exercise 1.9.
6 Sports Math

K’s with Kershaw: Terminal Velocity and Drag Forces

Our first three examples all utilized an assumption that gravity was the
only force to be considered. Such an assumption is valid on the Moon, but is
not realistic in any sports situation. So, why would we consider such exam-
ples? The only reason is mathematical convenience. While the mathematics in
Example 1.3 may have become uncomfortably detailed, the underlying equa-
tions in Table 1.1 are about as simple as equations get. For a vertical jump
and a basketball shot, we can hope that other forces do not have a very large
effect, but for many sports this hope is in vain. In this section, we add air re-
sistance to the mix. The good news is that the mathematical model becomes
more accurate; the bad news is that the resulting equations are, for the most
part, unsolvable.
Let’s start with a thought experiment. If you hold your hand out of the
window of a moving car, how much force will you feel? The faster the car is
going, the more force, right? Also, you can increase or decrease the force by
changing your hand position: more hand facing the front of the car means
more force. These illustrate the main principles of air drag. The magnitude
of the force depends on the speed of the object. In most sports situations, a
good approximation is that the force due to drag is

Fd = cv 2

where Fd is the magnitude of the drag force, and v is the speed of the object.
The scalar c depends on such influences as humidity, temperature, altitude, the
composition and orientation of the object, and other factors. One of the other
factors is the speed of the ball, but to keep things simple we will ignore this.
Most of the influences are well known to us: at high altitude the air is thinner
and air drag is reduced: the air drag at elevation 7350 feet (e.g., Mexico City)
is 23% less than that at sea level. Under most conditions, higher temperature
reduces air drag. The main reason that balls fly farther in warm weather is
humidity: contrary to most people’s intuition, higher humidity causes lower
air drag.
For a falling object, Newton’s Second Law now looks like

a = −g + kv 2

where k equals the c from the drag force divided by the object’s mass m.
Note the signs: gravity pulls in the negative direction while (since the object
is falling) air drag pushes upward. Let’s track acceleration in another thought
experiment: if you jump out of an airplane, what will happen to your velocity?
At first, kv 2 is smaller than g so that a is negative. A negative change in
velocity means that your downward speed will increase. But, the faster you
fall the more air drag you experience and your speed gradually approaches an
 Projectile Motion 7

equilibrium value at which the air resistance kv 2 exactly balances gravity g.

At this point, your acceleration is zero and your velocity will remain constant.
(This is Newton’s First Law: an object maintains a constant velocity unless
acted on by an external force.) You will continue to fall at this speed unless
there’s a change in force: e.g., your parachute opens, or you change body
orientation to alter air resistance. If your parachute does not open, the name
terminal velocity for this speed is appropriate, if harsh.

Example 1.4 If the terminal velocity of a baseball is 95 mph, find the value
of k for air drag on the baseball.
Solution. One of our values for g is 32 ft/s2 , so we need to convert mph to
ft/s. Given 5280 feet in a mile and 3600 seconds in an hour, we have
95 mi mi ft 1 hr ft
hr = 95 hr · 5280 mi · 3600 s ≈ 139.3 s . At terminal velocity, a = 0 so
g 32 ft/s2 −1
−g + kv 2 = 0. Then k = 2 = 2 ft2 /s2 ≈ 0.00165 ft .
v (139.3)

We can now estimate how many miles per hour a Clayton Kershaw fastball
loses as it travels to home plate. To do this, we use the above value for k and
assume that a pitch travels horizontally. We also assume that the pitch travels
55 feet (60 feet, 6 inches minus a long stride and stretch).

Example 1.5 For a baseball that starts at 95 mph and travels horizontally
(no gravity), find its speed after it travels 55 feet.
Solution. In the absence of gravity, Newton’s Second Law is a = −0.00165v 2 .
This assumes that the pitch moves in the positive direction, meaning that the
air drag is in the negative direction. Unfortunately, this equation requires
calculus to solve. As seen in the calculus box that follows, the solution to
v0
the general equation a = −kv 2 gives velocity v = and position
1 + kv0 t
1
x = ln (1 + kv0 t). We first need to know how long it takes for the ball to
k
travel 55 feet; we can then plug this time into our velocity equation. To get
1
55 = ln (1 + kv0 t), we need 55k = ln (1 + kv0 t) or 1 + kv0 t = e55k . Then
k
kv0 t = e55k − 1 or t = (e55k − 1)/(kv0 ). With k = 0.00165 ft−1 and v0 = 139.3
v0
ft/s, we get t ≈ 0.413 s. At this time, the velocity is ≈ 127.2 ft/s or
1 + kv0 t
86.7 mph. Kershaw’s fastball loses 8.3 mph, or just under 9% of its speed.

One consequence of this loss in speed has to do with radar guns. Clearly,
the point at which the pitch speed is measured can drastically change the
reading on the radar gun.

Calculus Box: Solving for Velocity

In Example 1.5, we needed to solve the equation a = −kv 2 . Given that
dv dv
a= , what we need to solve is the differential equation = −kv 2 . This
dt dt
8 Sports Math

is a separable differential equation; its solution technique is covered in many

calculus books. (See Smith and Minton.) As the name implies, we want to
1 dv
separate the variables v and t, so we rewrite the equation as − 2 = k and
v dt
then integrate.
Z Z
1 dv
− 2 dt = kdt
v dt
1
= kt + c
v
1
v=
kt + c
where c is a constant to be evaluated using the initial condition v(0) = v0 .
1 1
With t = 0 and v = v0 our equation becomes v0 = , and we get c = .
c v0
Substituting this in, we get
1 v0
v= =
kt + 1/v0 kv0 t + 1

as stated in Example 1.5. To get the position function, we use the fact that
dx
v= and so
dt
Z Z Z
v0 1 kv0 1
x= vdt = dt = dt = ln |kv0 t + 1| + c
kv0 t + 1 k kv0 t + 1 k

where c is again a constant to be evaluated. Since x(0) = 0, we have 0 =

1
ln(1) + c = c and so c = 0. We are left with x = ln (1 + kv0 t) as desired.
k

Bending with Bubba: Magnus Force

Bubba Watson was in a playoff for the 2012 Masters golf tournament, 163
yards from the hole and aiming about 45 yards left of the green because of
trees. His wedge hooked onto the green about 10 feet from the hole, setting
up his playoff victory. United States midfielder Jermaine Jones scored an out-
standing goal in the 2014 World Cup, bending a ball around a defender into
the net. Without the curve on the ball, the shot would have sailed several
paces wide of the goal. How did Watson and Jones do it? The role of spin in
causing ball trajectories to be curved is examined next.
The spin on a ball in flight is likely to create a force, called the Magnus
force, that causes the familiar curves of golf balls, soccer/footballs, baseballs,
and tennis balls. A basic understanding of the geometry of this force provides
 Projectile Motion 9

insight into the techniques used to control the flights of the balls. Trajectories
of balls acted on by gravity, drag, and the Magnus force must be simulated
on a computer. We focus on the geometry of the force here.
The Magnus force can be explained by New-
ton’s Third Law: every action has an equal
and opposite reaction. Imagine a ball spinning
clockwise with air flowing left to right, as in the
figure. The motion of the ball pulls the flowing
air downward. When the ball pulls the air down,
FIGURE 1.3: Air Flow
the air will push the ball up. This upward force
is the Magnus force. This explanation makes it
clear that the greater the spin rate of the ball, the greater the Magnus force.
The Magnus force depends on both the ve-
locity of the ball and its spin vector, defined as
follows. Consider a ball with a horizontal ve-
locity directly away from you. As you look at
the ball, the back of the ball is rotating from
top to bottom as shown in Figure 1.4. This is FIGURE 1.4: Spin
called backspin and is the most common spin
in sports. We define a spin vector s by first iden-
tifying the axis about which the ball rotates. In this example, the spin axis is a
left-right horizontal axis through the center of the ball. Now, imagine curling
the fingers of your right hand around the front of the ball with your fingers
pointing in the direction the ball is spinning. If you extend your thumb, it
should point to the right, which is the direction of the spin vector s.
The direction of the Magnus force is perpen-
dicular to both the spin vector (to the right) and
the velocity vector (pointing away from you).
The directions “up” and “down” both qualify;
we can determine which one is correct using an-
other right-hand rule. Point the index finger of
your right hand in the direction of the spin vec-
tor, holding your hand in such a way that you
can curl your bottom three fingers toward the FIGURE 1.5: Right Hand
velocity vector. Your thumb will point in the di-
rection of the Magnus force; in this case, up. (In calculus terms, if s is the spin
vector and v is the velocity vector, the Magnus force is in the same direction
as the cross product s × v.)
10 Sports Math

The first brief take-away is that backspin

creates a Magnus force with an upward com-
ponent. In Figure 1.6, the velocity vector (la-
beled “Vector”) is horizontal and to the right.
In this orientation, backspin means that the ball
rotates counterclockwise. The Magnus force is
straight up.
Because of the upward Magnus force, the ve-
FIGURE 1.6: Upward
locity will not remain horizontal, but will gain a
vertical component. The Magnus force and ve-
locity vector remain perpendicular throughout
the ball flight, so a change in velocity direction
is accompanied by a change in Magnus force di-
rection. Tilt the velocity vector as the ball goes
up and the Magnus force tilts backwards as in
Figure 1.7.
As the ball comes down, the Magnus force
tilts forward as in Figure 1.8. The Magnus force
is now pointing up and forward, giving the ball FIGURE 1.7: Backward
a positive horizontal acceleration. This has im-
plications in the catching of high pop-ups in
baseball, as we explore in exercise 1.27.
Figures 1.6-8 can also help us understand
what happens if the spin vector has a side-
ways component. You will need to mentally sub-
stitute the spin vector for the velocity vector
shown in the Figures, and also assume that the
velocity vector is into the page.
So, how do Bubba Watson and Jermaine
Jones bend their shots? Pure backspin can be
created if Watson’s golf club (or Jones’s foot)
slides directly under the ball. Because the club FIGURE 1.8: Forward
is coming downward and hitting the ball below
center, during the brief time that club and ball
are in contact the ball will start to roll up the club; this creates the ball’s spin.
With pure backspin, we have seen that the Magnus force will be upward (and
backward or forward depending on whether the ball is going up or down).
If, however, the clubhead is tilted or the swing path is crooked, the ball
will not roll straight up (vertically), but instead will roll some to the side;
the ball has some sidespin. Now, assume that the balls in Figures 1.6-8 have
velocity into the page and “Vector” shows the direction of the spin vector.
Watson swings so that at impact the club slides under the ball and is moving
from right to left. The ball rolls up the club and to the right, producing a spin
vector like that of Figure 1.8. The resulting Magnus force will bend the ball to
the right, as we saw at the Masters. ESPN’s Sport Science estimates the axis
 Projectile Motion 11

of rotation was tilted by 38 degrees, creating the severe rightward curve that
he needed. Jermaine Jones, kicking with his right foot, created a spin vector
like that of Figure 1.7, and his shot curved to the left.

Example 1.6 For the following sports situations, identify a possible spin
direction, Magnus force direction, and resulting curve of the ball. (a) tennis
serve, (b) baseball fastball, (c) baseball curveball, (d) golf outside/in swing.
Solution. Answers will be given for right-handed players. (a) Spin on a tennis
serve can come from the motion of the racket and from the ball toss. Imagine
a ball dropping from a large height and then contacting a racket squarely.
Because the ball is moving downward, as it contacts the strings of the racket
it will roll down the racket. This is the exact opposite of the spin in Figure 1.4
(and is called topspin). The spin vector will be to the left and the Magnus
force downward (for a serve moving horizontally). A right-handed player can
easily cause the racket to be moving from left to right as it strikes the ball.
This tilts the spin axis as in Figure 1.7 and the ball will slice to the left.
(b) A fastball is released by letting the ball roll off of the index and middle
fingers. The spin is backspin as in Figure 1.6, with an upward Magnus force.
A right-handed pitcher’s arm is likely to not move in a purely “overhanded”
fashion, but have some right to left movement. This will cause the two fingers
to be tilted to the right, creating a spin as in Figure 1.8 (spin vector down and
right), and a Magnus force up and right. This creates some movement to the
right, as in the “tailing fastball” made famous by Greg Maddux and others.
(c) The curveball is released with a snap of the wrist that causes the back of
the ball to rotate upward (topspin again) and (for a righthander) to the right.
This makes the spin vector point up and to the left, with a Magnus force
that is down and to the left. The curveball will move a small amount to the
left, but its main movement is downward. (d) A good golf swing always goes
underneath the ball, so there is always backspin. For a righthanded golfer,
outside/in means that the club is moving somewhat from the right to the left
as it moves into the ball. The result is seen in Figure 1.8, with spin vector
down and right and Magnus right up and right. This is the classic slice that
plagues so many golfers.

Smiling with Dimples

Golf balls have 300-400 dimples, small in-
dentations that cover the ball in a symmetric
pattern (most asymmetric patterns are illegal,
as an asymmetry may prevent the ball from
curving). Historically, golfers had noticed that
golf balls flew farther and straighter after get-
12 Sports Math

ting battered and dented. Golf ball manufacturers stepped in and offered “pre-
dented” balls that revolutionized the game. The advantages of dimpled golf
balls can be stated in terms of the forces on the ball. A dimpled ball traveling
150 mph with 3000 rpm of backspin has less than half the air drag of a smooth
ball; this may seem counterintuitive, but the roughness of the dimpled ball
creates a turbulent wake that results in a shorter separation of streamlines
and less drag. The television show Mythbusters created a dimpled car that
got demonstrably better gas mileage.
But, that is not the only advantage of the dimpled ball. At 150 mph and
3000 rpm, the dimpled ball has four times the lift force (upward Magnus force)
of a smooth ball. The increased Magnus force lifts the ball higher, giving it
time to travel farther, and providing spin to create consistent and controllable
trajectories. New golf ball designs go through extensive aerodynamic testing
to create ideal drag and lift profiles.

Calculus Box: A General Model of a Ball in Flight

If you mathematically model the forces of gravity, air drag, and Magnus
force, you have a highly accurate equation for the flight of a ball. You also
have an equation that cannot be solved (meaning that a solution cannot be
found in terms of the basic elementary functions that you commonly see in
mathematics courses). In our digital age, this is not a deal-killer. We can throw
the equations into Mathematica or other software and get a nice graphical
“solution” that gives us important information.
The most common form for air drag is −kv 2 where v is the speed of the
ball and the minus sign indicates that the force is in the direction opposite
that of velocity. In vector form, this becomes −kd |v| v where v is the velocity
vector. The Magnus force is generally given as km s × v where s is the spin
vector described above. A large complication that is disguised by these ex-
pressions is that kd and km are, in general, not constants. They can depend
on speed, spin rate, and various environmental factors. The drag coefficient
kd , for example, may depend on speed in a complicated fashion, where there
is a sudden decrease in drag at a critical speed. Even treating kd and km as
constants, the resulting differential equation
dv
= −g − kd |v| v + km s × v
dt
is highly coupled and therefore difficult to solve. Here, “coupled” means that
the x-, y-, and z-dimensions do not split off into separate equations. The actual
components
q for a baseball can be found in exercise 1.45. For now, notice that
|v| = vx + vy2 + vz2 and the x-component of s × v is sy vz − sz vy so that you
2

must know vy and vz to solve for vx (and you need vx and vz to solve for vy ).
 Projectile Motion 13

The Effects of Drag and Lift

There are numerous software packages that can give you nice graphs repre-
senting solutions of the above equations. Below, we look at graphs of baseballs
in flight to get an idea of the importance of drag and lift (Magnus force) in
baseball.
Before doing so, we want to get an idea of what the software is doing
to create these graphs. Having some idea of what is going on “behind the
curtain” helps us to understand the limitations of the software and be more
intelligent users of the technology. Plus, it is fun to know what’s going on!
This is a fully three-dimensional model of motion. Let’s start by orienting
ourselves. For baseball, one choice is to have the batter at the origin, the y-
axis horizontal through the batter and the pitcher, the x-axis horizontal at a
right angle to the y-axis, and the z-axis vertical. We use the general equations
from above with appropriate values for the coefficients of kd = 0.0018 and
km = 0.000064 (which we get from Watts and Bahill). To make life a little
simpler, let’s focus on the equation for acceleration in the z-direction.
q
az = −32 − 0.0018vz vx2 + vy2 + vz2 + 0.000064(sx vy − sy vx )

where sx is the x-component of the spin vector in units of rpm.

For initial values of the variables, let’s consider a ball hit directly to center
field at a height of 3 ft with initial speed of 160 ft/s at an angle of 30 deg
above horizontal, and with backspin of 4000 rpm. Then the initial position is
x(0) = 0 ft, y(0) = 0 ft, and z(0) = 3 ft. The initial velocity is vx (0) = 0 ft/s
(the ball is hit directly toward center field), vy (0) = 160 cos(30◦ ) ≈ 138.6 ft/s,
and vz (0) = 160 sin(30◦ ) = 80 ft/s. The initial values for spin are sx = 4000
rpm, sy = 0 rpm, and sz = 0 rpm (if the spin is pure backspin, the spin vector
points directly to the right, the positive x-direction).
Let’s substitute these values in and compute the initial acceleration in the
z-direction. Note that the square root term represents speed, so instead of
trying to compute the term as listed we can simply substitute in the initial
speed of 160 ft/s. We have −32 ft/s2 from gravity, −.0018(80)(160)≈ 23 ft/s2
from drag, and approximately .000064(4000)(138.6) ≈ 35.5 ft/s2 from lift.
Notice that at the beginning of the flight of the baseball, lift from the
Magnus force is actually larger than the gravitational pull. The drag force is
more than 70% of the force due to gravity. In many sports situations, it is not
at all reasonable to ignore drag or lift.
Back to the calculation: if we add the pieces together, we get an accelera-
tion in the z-direction of about −19.5 ft/s2 . Because this acceleration is not
constant, we cannot use it to compute velocity exactly. However, if the accel-
eration is approximately constant for a tenth of a second, then the change in
velocity for that tenth of a second is approximately acceleration times change
14 Sports Math

in time and
vz (0.1) ≈ vz (0) + 0.1az (0)
from Table 1.1. Substituting in, we get vz (0.1) ≈ 80 ft/s + (0.1 s)(−19.5
ft/s2 ) = 78.05 ft/s. Given a z-velocity, we can estimate the z-position from
z(0.1) ≈ z(0) + 0.1vz (0.1) and get z(0.1) ≈ 3 ft + (0.1s)(78.05 ft/s)≈ 37.8 ft.
There are numerous details to which you can reasonably object. Before
doing so, let’s review the general procedure just outlined. We pick a time step
dt (we used dt = 0.1 above) and then
1. Start with the current values for velocity and position.
2. Compute the acceleration.
3. Use acceleration to update velocity (as if acceleration were constant).
4. Use velocity to update position (as if velocity were constant).
Now, to address some objections that you may have. The assumptions that
acceleration and velocity are nearly constant may not be acceptable (in fact,
in our calculation above, velocity changed from 80 ft/s to about 78 ft/s). If we
decrease the value of dt, however, the assumption should be more accurate.
Theoretically, we can decrease dt until we are happy with the assumption. In
practice, the smaller you make dt the more calculations have to be made to
reach a target value of t (for example, if the flight time of our baseball is 4
seconds, with dt = 0.1 we need 40 updates to reach t = 4; with a smaller
dt = 0.01 we would need 400 updates). In our sample calculation, we used
the velocity update of 78.05 ft/s to estimate the height. We could have used
the initial velocity of 80 ft/s or (even better) an average of the two. These are
issues that you can explore in a course in the field of numerical analysis.
The following graph shows the flight of the baseball described above under
three assumptions.

FIGURE 1.9: Effect of Drag and Magnus Forces

The graph labeled “gravity” shows the trajectory of a ball acted upon by
gravity only (no drag, no lift). This ball is launched at 160 ft/s (about 109
mph) at an angle of 30 deg. Its range is about 700 ft. How much effect does
drag have? The graph labeled “+drag” shows the trajectory of a ball acted
upon by gravity and air drag. Its range has been reduced by almost half, down
to 385 ft. And what about the Magnus force? The graph labeled “+Magnus”
shows the trajectory of a ball acted upon by gravity, air drag, and the Magnus
force. The lift force restores some distance, boosting a 385-foot out to a 460-
foot home run. The times of flight for the three hits are 5 s, 4.2 s, and 6.8 s,
 Projectile Motion 15

respectively. The backspin on the hit gives the ball extra height, extra hang
time, and extra distance.
A final note on the graphs in Figure 1.9 has to do with symmetry. The
“gravity” graph has symmetry; the flight up is the same as the flight down. In
the “+drag” graph, you may notice that the flight down is shorter than the
flight up; since drag is reducing the speed, less distance is covered later in the
flight. The “+Magnus” graph is also asymmetric, with the peak of the graph
occurring at about the 60% mark of horizontal distance.
Figure 1.10 shows a side view of three baseball pitches, to illustrate the ef-
fect of spin. All three graphs show 90-mph pitches. The middle graph includes
the effects of gravity and air drag, but no Magnus force. The top graph shows
the lift created by 2400 rpm of backspin on a fastball, while the lowest graph
illustrates a curve ball with 2400 rpm of topspin.

FIGURE 1.10: Effect of Spin

The middle graph is included for reference. As we see in the next section,
it does not represent a baseball pitch thrown with no spin.

Knuckling Down
The knuckleball is a baseball pitch known for its contrariness. Its motion
fools batters, catchers (Bob Uecker famously said that his technique for catch-
ing the knuckleball was to wait for it to stop rolling and then pick it up), and
umpires. Even its name is contrary: the ball is not gripped with the knuckles,
but with the fingertips (making the knuckles visible to the batter). Knuckle-
balls are not thrown hard (60-80 mph), have very little spin, and are said to
dance around: the web site FanGraphs has several excellent video clips of R.A.
16 Sports Math

Dickey’s knuckleball moving first one way and then the other. The explanation
of how the knuckleball works is also unusual.
A baseball has raised stitches holding its
cover together. The stitches are raised enough
to create a lift force, much like the curved wing
of an airplane. Watts and Bahill took wind tun-
nel measurements of the force for different ori-
entations of the ball. If θ measures the angle of
rotation of the ball from some initial orienta-
tion, the lateral force on the ball is approximately −0.1 sin(4θ) lb. (Note that
θ = 2π ≈ 6.3 radians represents one rotation of the ball.)
As the ball rotates, the lateral force on the ball due to the stitches oscillates.
In this case, lateral force means side-to-side or left-to-right from the batter’s
perspective. At ball orientation θ = 0 there is no force on the ball. As it
rotates to positive values of θ, the lateral force goes negative and the ball
will accelerate to the left. As θ increases past θ = π4 , however, the lateral
force turns positive. The ball now has an acceleration to the right; the ball’s
leftward movement slows, and the ball may even start moving to the right (as
seen at FanGraphs). Following the force in this way can be awkward. As seen
in the calculus box below, we can solve for the lateral (left-right) position of
the ball and learn more.
Figure 1.11 (not to scale) shows the path
of a knuckleball with almost no spin (ω = 0.1
rad/s). This is a blimp’s-eye view of the pitch
from above, with the pitcher on the horizontal
axis to the far left and home plate on the hor-
izontal axis to the far right. The ball finishes FIGURE 1.11: ω = 0.1
about 2 inches to the left of home plate. So,
this pitch behaves like a 65 mph curveball without much break.
For Figure 1.12, the spin rate is increased
to ω = 2 rad/s (about a quarter turn of the
ball from pitcher to home plate over 0.65 sec-
onds) and the trajectory changes dramatically.
The vertical scale is in feet, so this pitch moved
0.3 feet or 3.6 inches to the left before turning
around and coming back to the center of the
FIGURE 1.12: ω = 2
plate.
For Figure 1.13, the spin rate is further in-
creased to ω = 10 rad/s (about one full rotation
of the ball) and a different picture emerges. It
is true that the ball is oscillating very rapidly,
but the movement is nearly imperceptible and
at 65 mph would be unlikely to confuse a major
league batter. FIGURE 1.13: ω = 10
We see that the knuckleball is extremely sen-
 Projectile Motion 17

sitive to the spin rate, with the best results occurring when the ball rotates
between one-quarter and one-half of a turn on the way to the plate. This gives
some idea of the difficulty of throwing the pitch effectively, and why so few
pitchers have had extended success with the knuckleball.

Calculus Box: Lateral Position of a Knuckleball

With spin rate ω rad/s, the lateral position x(t) of a pitch thrown directly
toward home plate obeys Newton’s Second Law

m x00 (t) = −0.1 sin 4θ

Tracking x in feet, we use a ball mass of m = 0.1 slug. If the ball spins at a
rate of ω rad/s, then 4θ = 4ωt + θ0 rad, for some initial orientation θ0 . Our
equation becomes
x00 (t) = −10 sin(4ωt + θ0 )
Integrate this with respect to t to find the lateral velocity. Thus
10
x0 (t) = cos(4ωt + θ0 ) + c
4ω
where the integration constant c can be determined from the initial condition
x0 (0) = 0 (the ball is thrown directly at home plate). We find c = − 4ω
10
cos θ0
and so
5 5
x0 (t) = cos(4ωt + θ0 ) − cos θ0
2ω 2ω
We integrate one more time with respect to t to get
 
5 5
x(t) = sin(4ωt + θ 0 ) − cos θ 0 t+c
8ω 2 2ω

where the integration constant c is determined from the initial condition

x(0) = 0. We find that c = − 8ω5 2 sin θ0 and conclude that
 
5 5
x(t) = [sin(4ωt + θ 0 ) − sin θ 0 ] − cos θ 0 t
8ω 2 2ω
for a general equation of the knuckleball. Figure 1.10 uses θ0 = 0 and Figures
1.11 and 1.12 use θ0 = π2 .
18 Sports Math

Exercises
In exercises 1.1-12, assume that gravity is the only force. T refers to think-
ing problems, conceptual problems requiring no calculations. C refers to
problems requiring calculus or significant computer calculations. P refers to
projects; these are ideas for further investigation (hints and resources are at
the book’s web site).

1.1 An object is dropped from a height of 100 m. Find its velocity and position
at times t = 1, t = 2, and at impact.
1.2 Find the impact speed for divers dropping from heights of (a) 32 ft and (b)
128 ft. (c) Fill in the blank, and prove that you are correct: If height is multiplied
by 4, impact speed is multiplied by .
1.3 An object falls from a height of 400 m. Find its position at times t = 1,
t = 2, t = 3, and t = 4. Compute dp1 = p(1) − p(0), dp2 = p(2) − p(1),
dp3 = p(3) − p(2), and dp4 = p(4) − p(3). Show that dp2 = 3dp1 , dp3 = 5dp1 , and
dp4 = 7dp1 . Conjecture values of dp5 and dp6 . To explain this pattern, compute
22 − 12 , 32 − 22 , 42 − 32 , and so on.
1.4 A diver jumps from a height of 10 m with an initial (upward) velocity of
1 m/s. Ignoring horizontal motion, (a) find the diver’s velocity and position at
times t = 1 and t = 2. (b) Find the diver’s maximum height, time to impact,
and velocity at impact.
1.5 Dwight Howard set a record by making a mark on a backboard 12’6” above
the ground. His vertical jump was measured at 39.5 inches. Find (a) the required
initial velocity in ft/s and (b) Howard’s hang time. (c) The episode “Flight” of
the television series Sport Science states that a hang time of one second is the
limit of human abilities. If this is true, what is a human’s maximum vertical
jump?
1.6 A football punt has a hang time of 5 seconds. Ignoring horizontal motion,
find (a) the required initial velocity in ft/s and (b) the maximum height of the
football.
1.7 In Example 1.3, a 25-foot shot is launched with initial speed 30 ft/s at an
angle of 50 deg. (a) Is the shot good if the angle is 45 deg? (b) Is the shot good
if the speed is 31 ft/s (angle of 50 deg)?
1.8 In Example 1.3, use trial and error to approximate to the nearest one-tenth
s1 and s2 , the smallest and largest speeds, respectively, for which the shot is
good.
1.9 Show that a 25-foot shot launched with initial speed 30 ft/s at an angle of
48.5 deg is good. Show that the shot reaches basket height at a shorter distance
for both angles of 48 and 49. Conclude that at speed 30 ft/s it is not possible to
hit the back rim. This is a characteristic of the launch speed that has the largest
margin of error in angle: it is the smallest launch speed to reach the center of the
basket. Explain why this could be called the “best” launch speed, and explain
how a shooter could “feel” this ideal speed.
 Projectile Motion 19

1.10 A shortstop throws to first base from 80 ft away, releasing the ball horizon-
tally from a height of 5 ft with speed 80 mph. (a) What happens to the throw?
(b) By trial and error, find the angle such that the ball reaches first base at
height 5 ft. At this angle, how far above first base is the ball aimed?
1.11 A tennis serve launches a ball with initial speed 120 mph from a height
of 10 feet at an angle of 7 deg below horizontal. To be good, the serve must
clear a 3-foot-high net that is 39 ft away, and must hit the ground on or before
the service line 60 ft away. (a) Is this serve good? (b) By trial and error, find
the angles such that the ball clips the top of the net and the ball lands on the
service line; the difference in angles is the margin of error at 120 mph. (c) Find
the margin of error of a 135 mph tennis serve. Does the extra speed increase or
decrease the margin of error?
1.12 A baseball is hit from a height of 3 ft with initial speed 130 ft/s at an angle
of 25 deg above the horizontal. Will the ball clear a 6-ft wall located 400 ft away?
1.13 A free kick is taken from 25 yd out from a spot even with the right post.
The initial speed is 90 ft/s and the velocity is 4 deg to the right of the post. The
kicker puts sidespin on the ball that results in a constant acceleration of 16 ft/s2
to the left. (a) If the goal is 24 feet wide, will this shot be on goal? (b) If there
is a wall 10 yd from the kicker whose rightmost defender extends one foot to the
right of the goal, does the kick hit the wall?
1.14 Suppose that a ball is designed with a 23% reduction in constant k from
the ball in Example 1.4. By how much is the terminal velocity changed?
1.15 The terminal velocity of a tennis ball is 70 mph. Find the constant k for its
drag, and estimate the speed of a 135 mph serve after it has traveled the 60 ft
to the service line.
1.16 The terminal velocity for a badminton shuttlecock is 6.8 m/s. Find the
constant k for its drag, and estimate the speed of a (world record) 493 km/hr
smash after it travels 10 m.
1.17 For the following sports situations, identify a possible spin direction, Mag-
nus force direction, and resulting curve of the ball. (a) tennis topspin shot, (b)
baseball sidearm fastball, (c) baseball fly ball to left field by a lefthanded batter,
(d) golf inside/out swing, (e) basketball free throw.
1.18 Here are the 10 m splits for Usain Bolt’s world record sprint of 9.58 s in
2009: (10, 1.89), (20, 2.88), (30, 3.78), (40, 4.64), (50, 5.47), (60, 6.29), (70, 7.10),
(80, 7.92), (90, 8.75), (100, 9.58). That is, he reached the 10 m mark in 1.89
seconds, the 20 m mark in 2.88 seconds, and so on. Use these to compute his
average velocities in each 10 m interval. Estimate his speed at 10 m, 20 m,
and so on. Which of these estimates is the least accurate? Why? Estimate his
acceleration in each 10 m interval.
1.19 At the 2014 Olympics, the winning time for the men’s 1500 m speed skating
race was 1:45.01. The winning time for the men’s 5000 m speed skating race was
6:10.76. Compute average speeds for each event. Explain why the speed for the
5000 m race is slower. Use these two measurements to predict the average speed
for a 3200 m race. The team pursuit is this length; the winning time was 3:37.71.
Compute the average speed for this race, and explain why the speed is faster
than your prediction. At the 2012 Olympics, winning times for the men’s 1500
m and 5000 m runs were 3:34.08 and 13:41.66, respectively. Compute average
speeds for these races. Explain in terms of forces why humans can skate faster
20 Sports Math

than they can run. The winning time in the men’s 4x400 m relay was 2:56.72.
Why is this faster than the winning 1500 m time?
1.20 (a) At the 2012 Olympics, the winning time for the men’s 100 m dash was
9.63 s. The winning time for the men’s 4x100 m relay was 36.84 s. Compute
average speeds for each event. Explain why the speed for the relay race is faster.
(b) By contrast, winning times for the 400 m and 4x400 m races were 43.94 s and
2 min, 56 s, respectively. Compute average speeds, and explain why the relay is
slower.
1.21 At the 2012 Olympics, the winning time for the men’s 1500 m race was 3:34.
The winning time for the men’s 4x400 m relay was 2:56. Compute average speeds
for each event. Explain why the speed for the relay race is so much faster.
1.22 Find world record times for different distances for men’s and women’s run-
ning races, and compute average speeds for each. Try to find a formula for a
function f such that f (d) is close to the average speed for a race of length d m.
1.23 A driver averages 140 mph for the first 10 miles of a race and 180 mph for
the second 10 miles of the race. What is the average speed for the first 20 miles
of the race?
1.24 The drag force is often written as 12 Acρv 2 where A is the projected surface
area of the object, c is a coefficient depending on numerous factors, ρ is the
density of the air, and v is the speed of the object. Tests of footballs have shown
that a punt moving in a spiral has a value of Ac that is about 18% of that of
an end-over-end kick. Given that air density in Denver is 18% less than in New
York, compare the drag on a spiral in Denver to an end-over-end kick in New
York.
1.25 Convert ω rad/s to rpm. If a pitch has 2000 rpm of spin and takes 0.45 s to
reach home plate, how many rotations does it complete?
1.26 A knuckleball is thrown with the ball gripped by the fingertips. Given that
a good knuckleball has little spin, why is this grip good? For pitches for which
more spin is better, explain how the ball can be gripped to increase spin.
1.27 T Figure 1.1 shows a graph of height h as a function of time t. If gravity
is the only force and the ball has a horizontal motion, explain why the graph of
h versus horizontal distance x would look the same.
1.28 T Look up world records for track races for men and women, and record
the year in which the world record was set. In cases where the record was set
more than, say, 5 years ago, speculate on why the best athletes in the world have
not broken the record recently.
1.29 T For a ball launched vertically with backspin, (a) describe the direction
of the Magnus force on the way up, at the top, and on the way down; (b) explain
why the ball cannot go up (or come down) in a purely vertical direction; (c)
explain why baseball infielders on high pop-ups often have to backpedal rapidly
to make the catch.
1.30 T Filip Bondy’s book The Pine Tar Game details a baseball game between
Kansas City and New York on July 24, 1983. George Brett, after hitting what
he thought was a go-ahead home run for Kansas City, was ruled out because the
sticky pine tar on his bat had spread too far up the handle. Suppose that the
ball hit the bat on one of the areas with pine tar. Explain how the ball could get
extra spin and travel farther because of the pine tar.
 Projectile Motion 21

1.31 T In Figure 1.4, the Magnus force direction turned out to be exactly
opposite the direction of the motion of the back of the ball (the arrows). This is
true in general. Illustrate this for two other arrow directions.
1.32 T A soccer player kicks the ball with the instep. For a rightfooted kick,
explain in terms of ball spin why the ball typically bends from right to left. If
the ball is struck with the outside of the foot, how does the spin change?
1.33 T Suppose a golfer takes swings that produce straight shots. If the
righthanded player’s left foot is moved closer to the ball (“closing” the stance),
explain why a hook (right-to-left movement of the ball) is likely to result.
1.34 T In 2014, the NCAA changed its baseball specifications to reduce the
size of the stitches (from being raised 1 mm to being raised 0.5 mm). The goal
was to produce more home runs. Explain in terms of drag force why this might
work.
1.35 T (a) The spin on a major league pitch can approach 2500 rpm. Explain
why the lift force from the stitches is not important. (b) Pitching in Denver (high
altitude), explain why you would expect the ball to reach home plate faster but
with less movement than at low altitude.
1.36 T A sprinter such as Usain Bolt quickly reaches a peak speed. If he main-
tains that speed, the forces on him must balance. The push of his feet against
the track provides a positive acceleration. Give two examples of negative forces
that cancel this out.
1.37 T In various types of racing, “drafting” is important. Explain in terms of
forces what advantage(s) drafting provides.
1.38 T In tennis, so-called “spaghetti” stringing styles (now banned) caused the
contact time between racket and ball to increase. Discuss the resulting changes
to spin rate, the Magnus force, and the playing of the game.
1.39 T If a baseball has a terminal velocity of 95 mph, explain why a pitch
thrown horizontally at 95 mph loses speed.
1.40 T The Puzzler’s Dilemma presents evidence that swimmers can swim as
fast in a thick, high-viscosity medium (e.g., syrup) as they swim in water. Clearly,
the drag in a thicker medium is higher than in a thinner medium. Explain what
force could compensate for the drag increase and allow swimmers to reach the
same velocity.
1.41 T Hold a (low-mass) racquetball directly on top of a (high-mass) soccer
ball, then drop the two balls simultaneously. The racquetball should bounce high
into the air. Use Newton’s Third Law to explain why the racquetball bounces so
much higher than the soccer ball.
1.42 T Imagine a golf putt that is uphill with no break. Suppose the putt starts
out slightly off line. Argue that gravity will pull the putt farther off line. Contrast
that to the effect of gravity on a downhill putt.
1.43 C For a constant acceleration a = c, derive the formulas for velocity and
position given in Table 1.1. Derive formulas for velocity and position for an
acceleration of a = ce−t .
1.44 C To find the angle of entry in Example 1.3, note that the slope of a curve
dy y 0 (t)
is given by = 0 . Find this slope at the entry time t = 1.29 and then
dx x (t)
22 Sports Math

use the fact that the absolute value of slope equals tanθ to find the entry angle
θ. To show that the ball does not touch the rim, we need to find the minimum
distance from the trajectory to the rim.p The distance between a point (x, y) and
the front of the rim at (24.25, 0) is d = (x − 24.25)2 + y 2 . We assume that the
ball follows the tangent line at the point of entry, which is y = −0.949(x − 24.9).
In the expression for d, replace y with −0.949(x − 24.9). Then solve the equation
d0 (x) = 0 to find the x-value of the closest point. Finally, show that the minimum
distance is greater than 4.75/12, the margin of error for the shot.
1.7 5
1.45 C The function f (ω) = − sin(2.72ω) gives the position in feet to
ω 8ω 2
the left or right of the center of home plate of a knuckleball when it reaches
home plate. Graph this function, and comment on the spin rate ω that produces
the best pitch. Find the value of ω that produces the maximum deviation. Find
the limit of the function as ω goes to 0.
1.46 C For a ball launched from the ground with initial speed v0 , show that
the (gravity-only) maximum range is obtained with a launch angle of 45 deg. If
the launch height and landing height are not the same, define α as the angle in
degrees above or below horizontal (e.g., the slope of the ground). Show that the
maximum range is obtained with a launch angle of 45 + 21 α.
1.47 C If drag is proportional to v instead of v 2 , rework Examples 1.4 and 1.5.
How much difference is there?
1.48 C Simulate baseball hits using equations vx0 (t) = −0.0018vx vx2 + vy2 + vz2 +
p

0.000064(sy vz − sz vy ), vy0 (t) = −0.0018vy vx2 + vy2 + vz2 + 0.000064(sz vx − sx vz )

p

and vz0 (t) = −32 − 0.0018vz vx2 + vy2 + vz2 + 0.000064(sx vy − sy vx ).

p

1.49 C Pitchers release the ball at different locations. The average release point
is about 55 feet from home plate. Assume that a pitcher releases the ball 53 feet
from home with an initial velocity of 93 mph. Compare the time it takes this
pitch to reach home to a 95 mph pitch released from the typical 55-foot mark.
1.50 P Find equations for tennis serves on the book’s web site. Explore the
effects of such variables as height of player; location of server along the baseline;
amount that server jumps into the court; serving up the middle versus wide;
amount of topspin on the ball. Howard Brody has published two excellent books
on the science of tennis.
1.51 P Find equations for golf shots on the book’s web site. Explore the effects
of such variables as amount of sidespin; changes in elevation; loft of the club.
Golf By the Numbers has more on these topics.
1.52 P For a track event such as the marathon, look at the changes in times
(world record times, or Olympics winning times, or other) over the years and
develop a formula to predict times in the future. You may need to take into
account changes in equipment and illegal doping practices. Try to determine the
limits of human performance. John Brenkus’s Perfection Point does this type of
investigation.
1.53 P Experiments to determine drag or lift forces are generally very difficult
to complete. However, here is a simple one you can complete with a simple
rangefinder that can compute speed. Take a coffee filter and let it drop. It should
quickly reach a terminal velocity that you can record. Then repeat the process
with a second coffee filter nested inside the first. All elements of the forces on the
 Projectile Motion 23

falling object have been maintained except that mass has doubled. If the drag
force is proportional to speed, then the terminal velocity should double. If the
drag force is proportional to√the square of the speed, then the terminal velocity
will increase by a factor of 2 ≈ 1.4.

Further Reading
Peter Brancazio’s Sport Science is an excellent introduction to the physics
of sports.
Robert Watts and Terry Bahill’s Keep Your Eye on the Ball is the best of
several physics of baseball books.
Equations of motion for baseball pitches were adapted from Alan Nathan’s
and Michael Richmond’s excellent web sites.
A search for “Physics of (favorite sport)” will bring up many results.
I can recommend The Physics of Basketball (Fontanella), The Physics of
Hockey (Hache), Football Physics (Gay), Newton’s Football (St. John and
Ramirez), The Science of Soccer (Wesson), The Physics and Technology of
Tennis (Brody, Cross, and Lindsey), Golf Science (Smith), The Science of
Golf (Wesson), Golf By the Numbers (Minton), and The Physics of NASCAR
(Leslie-Pelecky).
The John Brenkus television series Sport Science explored basketball play-
ers’ phenomenal jumping abilities in episodes called “Flight” and “Skywalk-
ing.” The Bubba Watson Masters tournament shot and Dwight Howard ver-
tical jump are featured in separate three-minute segments for ESPN.
Data about drag and lift forces for dimpled versus smooth balls can be
found in Golf By the Numbers and the Titleist web site.
Information about speed skating can be found in Gliding For Gold by
Denny.
An analysis of times for Usain Bolt’s races can be found at
http://sportsscientists.com/2009/08/analysis-of-bolts-9-58-wr accessed 7-21-
2015.
An article on the NCAA baseball stitches is at http://www.baseballamerica.com
/college/ncaa-to-switch-to-flat-seamed-balls-in-2015 accessed 7-21-2015.
The coffee filter experiment is from The Dick and Rae Physics Demo Note-
book.
Brody’s other tennis science book is Tennis Science for Tennis Players.
Further suggestions can be found in the notes at the Sports Math web site.
Chapter 2
Rotational Motion

Introduction
Roger Federer’s serve is poetry
in motion, a graceful knee bend and
arch of the back followed by a power-
ful overhead smash of the ball. His el-
egance belies the difficulty of the tra-
ditional service motion. Both arms
raise into the air, one tossing the ball
and the other cocking the racket be-
hind the head. As the racket unwinds
behind the head, the back foot is
brought near the service line. Then
both feet push the body up and into
the court as the racket meets the
ball at the top of the swing. Begin-
ners sometimes take shortcuts, such
as simply raising the racket up and
swinging it forward without cock-
ing it behind the head. Given that
all professional tennis players use
(roughly) the same technique, there
must be advantages to the awkward
manipulations of the serve. We ex-
plore the physics behind the tennis
serve as we look at rotational motion.
Along the way, we will also answer the following questions. Why are back-
swings in golf, tennis, and baseball pitching important? Why are tennis rackets
and golf clubs so large? Do tall golfers have an advantage? What is the opti-
mal length of a baseball bat? How do ice skaters perform fast spins? Why do
acrobats pull their bodies into those awkward-looking tuck positions?

25
26 Sports Math

Going in Circles
An object that is spinning in place has rotational
motion, but none of the translational movement dis-
cussed in Chapter 1. A spinning ball in flight and a
basketball player spinning to the basket have both
types of motion. To analyze the rotational motion,
we track its position using an angle θ (the Greek let-
ter theta). Suppose an object is traveling in a circle. FIGURE 2.1: Angle θ
Place imaginary x- and y-axes with the origin at the center of the circle. The
line segment from the origin to the object makes an angle θ with the positive
x-axis. By tradition, we measure θ in radians, with counterclockwise being the
positive direction.

Example 2.1 An object rotates counterclockwise at a constant speed, cross-

ing the positive x-axis at time t = 0 and completing one lap every 20 seconds.
Find an equation for its angle θ(t) for any time t.
Solution. With t in seconds, we know that θ(0) = 0, θ(20) = 2π (one lap is
2π radians), θ(40) = 4π, and so on. Assuming that θ(t) is a linear function,
2π − 0 π
we need a line through (0, 0) and (20, 2π). The slope is = and so
20 − 0 10
π
θ(t) = t.
10

In Example 2.1, the slope corresponds to rotational speed, more properly

called angular velocity (denoted by ω, the Greek letter omega). Then an-
gular velocity is the rate of change of the angular position θ. Completing the
trilogy, we define angular acceleration α (the Greek letter alpha) as the
rate of change of angular velocity. Table 1.1 for translational motion has its
rotational equivalent in Table 2.1.
TABLE 2.1: Formulas for Constant Acceleration
angular acceleration α=c
angular velocity ω = ct + ω0
angular position θ = 12 ct2 + ω0 t + θ0

In Example 2.1, the object has constant angular velocity, its angular ac-
celeration is zero, and Table 2.1 shows that θ is a linear function.
There is a simple connection between transla- TABLE 2.2: v to ω
tional variables and rotational variables. If an object
x = rcosθ
travels in a (two-dimensional) circle centered at the
y = rsinθ
origin with radius r, then the relationships in Table
|v| = r|ω|
2.2 hold.
That is, the translational speed of the object is the rotational speed times
 Rotational Motion 27

the radius of motion. We can use this fact and Table 2.1 to draw a simple
inference about tennis serves, golf swings, baseball pitches, and the like.

Example 2.2 Three tennis players start swings with zero angular velocity,
applying a constant angular acceleration c rad/s2 . Players A and C use rackets
of length L ft while player B’s racket is 10% longer. Players A and B rotate
through an angle of π radians while player C rotates through an angle that is
10% longer. Compare the final speeds of the ends of the rackets of the three
players.
Solution. For player A, θ = 12 ct2 rad and ω = ct rad/s. To rotate the racket π
q
radians, it takes time t such that π = 21 ct2 or t = 2π
c s. At this time, angular
√ √
velocity is ct = 2πc rad/s, and the racket speed is L 2πc ft/s. The same
calculations hold for player B except that√L is replaced by 1.1L (the length L
plus 10% of L). The racket speed is 1.1L 2πc ft/s, which is 10% larger than
that of playerqA. For player C, the final angle is not π but 1.1π. This occurs
2.2π
√
at time t = c s, and results in a racket speed of L 2.2πc ft/s. Since
√ √
2.2 ≈ 1.0488 2, this speed is about 5% higher than that of player A.

Example 2.2 shows that, all things being equal, (1) the longer the racket (or
golf club, or baseball bat) the higher the racket speed, and (2) the longer the
arc of the swing, the higher the speed. In this case, all things being equal means
that the angular acceleration remains constant. Unfortunately, sports swings
do not have constant acceleration throughout the entire arc of the swing. More
importantly, a longer racket is likely to be harder to swing, which probably
reduces angular acceleration (see Example 2.4 below). Given these caveats,
what can we learn about sports techniques?
The rotation of the tennis racket behind the
server’s head lengthens the arc of the service mo-
tion, thereby allowing for more speed to be gener-
ated. The ball is contacted with the server’s arm
and racket fully extended to maximize the length
L; the angular acceleration is aided by the server
jumping up a small amount. The tennis serve is
designed to maximize ball speed.
Golfers competing in Long Drive champi-
onships use preposterously long clubs, the drivers
reaching 48” long. The extra length provides ex-
tra distance, as in Example 2.2; this comes at
the expense of control, but long drive champions
must take that risk. Further, long drive champi-
ons usually take very big swings (well “past par-
allel” in golf jargon, with arms extended upward)
to increase the length of their swing arc.
28 Sports Math

Torquing Off Newton

Example 2.2 shows some of the effects of angular acceleration, but how
is angular acceleration produced? In Chapter 1, we saw that force is directly
related to translational acceleration. We multiply by the radius of the rota-
tional motion to translate rotational variables into translational variables, so
it should not be a surprise that we want to look at radius times force. We call
this quantity torque. In simple terms, torque τ (the Greek letter tau) equals
force times distance. Think of a wrench: you apply a force of F lb at the end
of a wrench of length L ft, producing a torque of τ = F L ft-lb that causes a
bolt to rotate.
Determining the distance L can be a little tricky. If you push on a door,
the torque depends on where and in which direction you push. Pushing on the
end of the door directly toward the hinges does not produce a torque (the door
won’t move); the relevant distance is not measured from hand to hinge, but
is the distance from the line of force to the hinge. If you are pushing toward
the hinge, the line of force goes through the hinge and the distance is zero.
No torque is produced, and therefore no rotation. You know by experience to
push on the end of the door at a right angle to the door; this maximizes the
distance of the line of force to the hinge, produces the maximum torque, and
closes the door most efficiently. (The vector definition τ = r × F, where F is
the force and r is the vector from the pivot point to the point of application
of the force, captures all of the details succinctly.)
Newton’s Second Law has the rotational equivalent

τ = Iα

relating angular acceleration to torque and I, the moment of inertia

(MOI). MOI is discussed in detail in the next section; first, let’s see a quick
example of torque in action.

Example 2.3 A football defensive player grabs a running back by the ankles,
applying a weak force of 80 lb for 0.1 s at a distance of 3 ft from the runner’s
center of mass. If the runner has a (head-over-heels) moment of inertia of 20
ft-lb-s2 , find the runner’s angular velocity.
Solution. We compute the torque using the formula τ = F L = (80 lb)(3 ft) =
240 ft-lb. We then have τ = Iα so that 240 ft-lb = (20 ft-lb-s2 )(α) or α = 12
rad/s2 . A constant angular acceleration of 12 rad/s2 for 0.1 s results in an
angular velocity of 1.2 rad/s. This is not a very fast rotation rate, but it does
mean that the running back’s body is rotating about his center of mass. He
will fall unless he can apply some sort of balancing force. Even a small force
can trip up a runner if applied far from the center of mass.
 Rotational Motion 29

All About MOI

An important realization is that the question “what is an object’s MOI”
is meaningless. In Example 2.3, the MOI was specified as a “head-over-heels
MOI” because a running back being tripped would tend to rotate head over
heels (or is that heels over head?). Just as mass can be defined as resistance to
translational acceleration, the moment of inertia can be defined as resistance to
angular acceleration. And there’s the problem: if the axis of rotation changes,
an object’s resistance to rotation is likely to change. It is much easier to spin
a person about a vertical axis than it is to rotate about the horizontal head-
over-heels axis. Thus, a person’s MOI about a vertical axis is smaller than
the person’s MOI about a horizontal axis. This is an important consideration
when thinking through rotations in gymnastics, diving, or skating. The lesson
is as follows. An object has a specific mass, but its MOI depends on
the orientation of the object and the axis of rotation.
A further distinction between mass and MOI can be illustrated by the
following experiment. Find a sledgehammer; it has a well defined (and large!)
mass. Grab it at the end of its handle and try to rotate it end over end (i.e.,
swing the sledgehammer); the resistance you feel is its large MOI for this
rotation. Now, grab the sledgehammer on its metallic hammer end and try to
rotate it end-over-end; this is a very easy rotation (small MOI). How does the
sledgehammer have different MOIs for two rotations that are both end-over-
end? In the first case, the heavy hammer end is a large distance away from
the rotation pivot; in the second case, the heavy hammer end is a very small
distance from the rotation pivot. So, MOI has to do with mass and the distance
that the mass is from the rotation pivot. If the sledgehammer is idealized to
a weightless handle and hammer mass concentrated at one point, then MOI
equals md2 , where m is the mass of the hammer, and d is the distance between
the hammer and the pivot. For sports, the following principle is helpful to keep
in mind.

The farther the mass is from the center of rotation,

the larger the MOI is for that rotation.

Example 2.4 Compare the MOIs and angular accelerations for the following:
(a) a mass of 2 kg rotating 3 m from the center of rotation; (b) a mass of 2
kg rotating 4 m from the center of rotation.
Solution. In this simple case, we can use the formula I = md2 to get (a)
I = 18 kg-m2 and (b) I = 32 kg-m2 . Moving the mass one meter (33%) farther
from the center increased the MOI by nearly 78%! The product of MOI and
angular acceleration is torque, so for equal torques the angular acceleration
18
in case (b) is 32 (about 56%) of the angular acceleration in case (a).
30 Sports Math

Size Is Important
Over the past thirty years, equipment in such sports as golf and tennis has
been completely transformed. In particular, golf clubs and tennis rackets are
much larger than in the past. In what way is bigger better?
Picking up the story in 1970, a retired aircraft engineer named Howard
Head developed tennis elbow. When he hit a tennis ball off-center, his racket
would twist and put pressure on the elbow. His engineering solution was to
reduce the twisting of the racket by increasing its MOI for off-center hits. The
calculation in Example 2.4 shows how to do it: move the mass farther from
the center. In this case, most of the mass of the racket head is in the frame,
so moving the frame away from the center by enlarging the racket head will
increase MOI. As an unexpected consequence, the wooden rackets of the day
could not be made with larger heads without becoming too heavy or cracking,
so Head designed a large, lightweight metal frame that launched the Prince
tennis racket revolution.

Example 2.5 Find the ratio of the MOIs for off-center hits on circular racket
frames of inner radii 12 cm and 13 cm, each of which is 1 cm thick.
Solution. The calculation in the calculus box below shows that the MOI is
a constant times f (R) = R3 + 32 R2 + R + 14 . Substitute in R1 = 12 to get
I1 = f (12), proportional to the MOI for the smaller racket. Substitute in
R2 = 13 to get I2 = f (13) for the larger racket. The ratio of the MOIs is given
by II21 ≈ 1.259. Therefore, enlarging a 12 cm radius frame by 1 cm (8.5%) to a
13 cm radius frame increases its MOI for off-center hits by 26%. Tennis rackets
are not circular, but a similarly large increase in MOI for tennis rackets was
obtained in the bigger rackets.

Calculus Box: Calculating MOI

We first approximate a tennis racket frame with
concentric circles of radii R and R+1 (assuming that
the frame is 1 cm thick). We also assume that the
frame has a constant mass density c kg/cm2 . The
rotation we are concerned with is a rotation about
the x-axis, as would occur if the racket were held
horizontally and the ball hit above or below the x-
axis. The moment of inertia
Z Z about the x-axis is given
by the double integral cy 2 dA where F is the frame of the racket.
F
This integral is best done in polar coordinates, where y = r sin θ. Then
 Rotational Motion 31

Z 2π Z R+1 Z 2π Z R+1
I= c(r sin θ)2 r dr dθ = c sin2 θ dθ r3 dr
0 R 0 R
1
(R + 1)4 − R4 = cπ(R3 + 23 R2 + R + 41 )

= cπ
4
for a circular racket frame. More realistic shapes produce more difficult inte-
grals.

Equipment Design
In the early days of metal frames on tennis rack-
ets, players would tape extra weight to the frame.
This weight, at a maximum distance from the cen-
ter of rotation of the racket, further increased the
MOI for off-center hits (and may have helped make
the racket swing more like a wooden racket). Rack-
ets featuring “perimeter weighting systems” (PWS)
began to appear, building extra weight into portions of the frame to improve
MOI and swing characteristics.
Similar concerns have impacted golf equipment design. Here, the concern
with twisting on off-center hits has less to do with pressure on the golfer’s
body and more to do with the consequences for the golf ball. If the golf club
twists, the initial direction of the ball will change, its spin will change, and
power will be lost. Golf equipment advertisements often throw around the
initials MOI. Which rotations are they trying to minimize?
Reading left-to-right, the picture to the right
shows a driver and two 3-woods. The 3-woods were
both good clubs for their days, which were 1995 (far
right) and 2010 (middle). The 3-wood grew a lot in
15 years! The 3-wood in the middle is about the size
of a 1995 driver, which is dwarfed by the modern
driver to its left. Why are larger clubs better? A
ball hit “on the toe” by one of these clubs contacts
the club outside of the center line perpendicular to
the club face. The club will rotate clockwise in re-
sponse. A larger MOI for this rotation reduces the amount of rotation, and
therefore produces a better outcome. The larger club has its mass at a larger
distance from the center of rotation, and therefore has the larger MOI for
this rotation. Golf clubs, like tennis rackets, grew larger to increase MOI and
improve performance. Drivers are now legislated to have a volume of 460 cc
or less.
32 Sports Math

Putters also changed over time. Examine the

two putters in the picture to the right. Which one
would perform better if the ball were struck above
or below center? Look at the backs of the putters,
where one is flat and the other curves upward
like a parabola. The putter to the right has extra
weight on its perimeter at a large distance from
the center, giving it a larger MOI for clockwise
or counterclockwise rotations caused by mishits.
The benefit to the athlete is in forgiveness. An imperfect swing that causes
a tennis or golf ball to be hit off-center will produce less rotation of the club,
with the resultant shot looking more like it would have if the ball had been
hit perfectly in the center of the club. One way of summarizing the previous
sentence is to say that the new rackets/clubs have “larger sweet spots.” This
manages to convey useful information without really making any sense. If
“sweet spot” refers to the best point on the racket/club to hit the ball, then
the sweet spot can move (a focus of modern golf club design) but you can’t
really make a point any larger. Nevertheless, the implication of the advertising
campaign is clear enough. The new rackets/clubs, by virtue of large MOIs and
other characteristics, have larger regions that will produce “good” shots that
are not much worse than your best. Since most of us spend most of our time
taking imperfect swings, this is good news.

Supercats and Tamedogs

Snowboarders, gymnasts, and divers all do acrobatics that involve rota-
tional motion. While the terminology of snowboarding can be more enter-
taining than enlightening (a supercat is a double backflip done on a straight
jump, and a tamedog is a frontflip performed on a straight jump), acrobatics
involve two main types of rotation. A rotation of the body about the verti-
cal axis (that is, rotating about an axis through the head and down between
the feet) is a “twist” and a rotation about a horizontal axis (that is, rotating
head over heels) is a “somersault” or “flip.” Somersaults, in turn, can be done
in different positions. The “tuck” has the body balled up with hands tightly
grabbing lower legs or ankles. The “pike” has the body bent in half at the
waist, with hands grabbing the back of the knees. The “layout” has the body
fully extended.
Can you order the MOIs for the tuck, pike, and layout? In the tuck posi-
tion, as much of the body as possible is pulled in close to the center of mass
(the center of rotation). While this is not an elegant position, it does min-
imize MOI and facilitate rapid rotations. The layout keeps as much of the
body as far from the center of mass as possible, maximizing the MOI and
 Rotational Motion 33

making multiple somersaults difficult. The pike position is in between. In The

Dynamics of Sports, Griffing gives typical values for a gymnast as 3.5 kg-m2
for the tuck position, 6.5 kg-m2 for the pike position, and 15.0 kg-m2 for the
layout position. Dives and gymnastic moves are given a degree of difficulty
rating based, in part, on the MOIs of the rotations.

Example 2.6 For a given torque, compare the times needed to complete a
triple somersault in the tuck position versus a double somersault in the pike
position.
Solution. Let the subscript tp refer to the tuck position and the subscript pp
refer to the pike position. Then since the torques are the same, τ = Itp αtp =
Ipp αpp or 3.5αtp = 6.5αpp where we use the MOIs from above. We get αpp =
3.5
6.5 αtp or αpp ≈ 0.54αtp . Assuming that the torque was applied for the same
amount of time t, then angular velocities are given by ωtp = αtp t and ωpp =
αpp t ≈ 0.54αtp t = 0.54ωtp . That is, the angular velocity in the pike position
is only slightly more than half the angular velocity in the tuck position, so
it takes nearly twice as long to complete a somersault in the pike position.
Thus, for a given torque, two somersaults in the pike position take almost as
long as four somersaults in the tuck position. Conversely, a triple somersault
in the tuck position takes about the same amount of time as 3(0.54) = 1.62
somersaults in the pike position. The double somersault in the pike position
is harder.

Keeping the Momentum

The concept of angular momentum gives us an easier way to do the
comparison in Example 2.6. Angular momentum L is defined by

L = Iω

and the torque equation τ = Iα is more generally interpreted as “torque equals

the rate of change of angular momentum” (since α is the rate of change of
ω). Thus, τ ≈ ∆L/∆t. Since the two divers in Example 2.6 apply the same
torque over the same time, their changes in angular momentum are the same,
and we can jump directly to Itp ωtp = Ipp ωpp or 3.5ωtp = 6.5ωpp .
A further consequence of τ ≈ ∆L/∆t is that when there is no torque,
angular momentum does not change. This is known as conservation of angular
momentum.

In the absence of torque, angular momentum remains constant.

We can now follow a diver or gymnast to a safe landing. While the tuck
position is used to complete several rotations, the landing is done in the layout
34 Sports Math

position (unless you’re doing a cannonball). What is the effect of the diver or
gymnast unfolding from the tuck position into the layout position? With the
athlete in the air, there is little torque being applied (perhaps a small amount
of air drag; let’s ignore that). If there is no torque, then angular momentum
is conserved. If ω1 is the angular velocity in the tuck position and ω2 is the
angular velocity in the layout position, then 3.5ω1 = 15.0ω2 or ω2 = 3.5 15 ω1 ≈
0.23ω1 . Just by changing positions in the air, the athlete’s rotation rate has
been reduced by 77%, which makes sticking the landing easier.
A flashy application of conservation of angular momentum is performed by
figure skaters. A figure skater starts a stationary spin with legs spread apart
and arms outstretched. Spinning about a vertical axis, the skater’s MOI is
relatively large. What happens if the arms are pulled tight to the body and
legs brought together? The skater’s actions reduce the MOI for the spin, and
the effect can be crowd-pleasing.

Example 2.7 A figure skater pulls in her arms and legs to reduce her MOI
for the spin by a factor of four. What is the effect?
Solution. By conservation of angular momentum, assuming no torques (so
ignoring air drag and ice friction) the product Iω remains constant. If the
subscript b denotes the beginning of the spin and the subscript n denotes the
new spin after the change in position, then Ib ωb = In ωn . We assume that
In = 41 Ib , so that Ib ωb = 14 Ib ωn from which we conclude that ωn = 4ωb . A
reduction of MOI by a factor of 4 results in an increase in spin rate by the
same factor of 4. The skater spins faster. The spin rate can be dramatically
manipulated by careful changes in MOI.

Conservation of angular momentum has one further consequence for spin-

ning. If you swing an arm in one direction, with no external torques, to con-
serve angular momentum your body will rotate in the opposite direction. The
magnitude of this movement is not large enough to impact diving or gymnas-
tics significantly, but falling cats use this fact to manipulate their bodies into
comfortable landing positions.

Exercises
In these exercises, T refers to thinking problems, conceptual problems re-
quiring no calculations. C refers to problems requiring calculus or significant
computer calculations. P refers to projects; these are ideas for further inves-
tigation (hints and resources are at the book’s web site).
2.1 An object starts at an angle of θ0 and completes one lap every T seconds.
Find a formula for the angle θ at time t if the object rotates (a) counterclockwise;
(b) clockwise.
 Rotational Motion 35

2.2 An object rotates at n Hz, meaning n laps per second. Compute the object’s
rotation rate in rpm and its frequency in rad/s.
2.3 Repeat Example 2.2 with a 20% increase in both length and angle.
2.4 (a) A racket is accelerated from rest with angular acceleration 4 rad/s2 . Find
the time needed to rotate an angle of π rad, and find the angular velocity at that
time. (b) Repeat part (a) with an angular acceleration of 5 rad/s2 . Compare the
percentage increases in angular acceleration and velocity.
2.5 Two softball pitchers create an angular acceleration of 15 rad/s2 on a ball
that is initially at rest. If one rotates the ball through an angle of π and the other
through an angle of 3π/2, compare the angular velocities of the balls. Speculate
on whether a constant angular acceleration throughout the pitching motion is
realistic.
2.6 (a) A BMX biker attempting a triple tail whip launches into the air at an
angle of 45 deg and speed 30 mph. Ignoring air resistance, how much time in the
air does the biker have? After accounting for air drag and launch and landing
preparation, suppose there is 1.2 s for the bike to complete its three rotations.
What is the average angular velocity during the spins? (b) To do a heelflip,
skateboarder Bucky Lasek must rotate his body 180 deg in 0.6 s. Find the average
angular velocity during the spin.
2.7 Repeat Example 2.3 with a larger force of 120 lb. Discuss the practical sig-
nificance of an increased angular velocity.
2.8 A golf club strikes a golf ball 0.5 in below its center of mass with a force of F
lb for 0.0005 s. Find the value of F needed to produce a spin rate of 7000 rpm.
2.9 In example 2.3, if the defender hits the runner in a line through the runner’s
center of mass, then there is no torque applied to the runner. In this case, we
compute linear momentum to see what happens. Linear momentum is given by
p = mv. If the runner has velocity 20 ft/s and mass 6 slugs, and the defender has
velocity −16 ft/s (the negative indicates a direction opposite that of the runner)
and mass 9 slugs, who has the larger momentum? The combined mass of the
runner/defender at contact is 15 slugs. Assuming that the total momentum after
the contact is the same as before, what is the velocity of the runner/defender
combination after contact? The runner has a higher speed, but the defender is
larger; in this case, who wins the battle?
2.10 In the situation of exercise 2.9, suppose that a defender has twice the mass of
the runner. How fast must the runner be moving to win the momentum contest?
Conservation of linear momentum, used in exercise 2.9, assumes that there are
no additional forces. Discuss forces that the runner and defender can apply to
overcome a deficit of linear momentum.
2.11 Compare the MOI and angular acceleration of a mass of 3 kg rotating 3 m
from the center of rotation to a mass of 2 kg rotating 4 m from the center. At
what distance would the 2 kg mass need to be to have the same MOI?
2.12 Repeat Example 2.5 with radii of 13 and 14 cm. Explain why the ratio is
different here than in Example 2.5.
2.13 A disadvantage of the larger tennis racket could be weight. Assuming a
constant mass density of ρ kg/cm2 , and using m = ρA where A is the area of the
racket frame, compare the masses of the rackets in Example 2.5. What would a
manufacturer need to do to have the two rackets have the same weight?
36 Sports Math

2.14 For a given torque, compare the times needed to complete three somersaults
in the pike position and two somersaults in the layout position.
2.15 A diver starts spinning in the layout position and then changes to the tuck
position. Compare the spin rates.
2.16 To move in a circle of radius r at constant speed v requires an acceleration
v2
(to change direction, not speed) of constant magnitude . The Indy car track
r
in Newton, Iowa, is 0.875 miles long per lap. If the track is circular and the
driver maintains a speed of 180 mph, find the magnitude of the acceleration
in the form 32c ft/s2 for some constant c. Since 32 is the magnitude of the
acceleration due to gravity, 32 is sometimes called 1 g. So c is the number of
g’s the driver experiences. (The real track is not circular, and is banked; sensors
show maximum g-forces of 5.5 at Newton.)
2.17 T Two runners have different builds but the same total mass. One has
slim lower legs and powerful upper legs, while the other has powerful lower legs
and slim upper legs. In terms of rotations about the leg socket, explain why the
first runner has the smaller MOI. In fact, long-distance runners tend to have
slim and relatively short lower legs.
2.18 T Over the years, elite golf and tennis players have become much taller.
In terms of club or racket speed, explain this trend.
2.19 T Discuss why there might not be constant angular acceleration in a golf
downswing. If constant angular acceleration is an invalid assumption, why would
we make it?
2.20 T Compare the MOIs of a football for a spiral and an end-over-end kick.
In 1900, a football was more like a modern rugby ball, much closer to spherical
than today’s football. Discuss how the modern ball makes passing easier (include
MOI as one aspect of your discussion).
2.21 T Golf ball manufacturers can alter the MOI of a golf ball. Spin is impor-
tant in golf to produce extra height and distance, and to have balls land softly,
but too much spin could cause a ball to go more up than out, losing distance.
Discuss advantages and disadvantages of having a higher MOI golf ball.
2.22 T The perimeter weighting system for a tennis racket (see page 31) puts
extra weight on the sides (long axis) of the racket. The racket is swung in an
upright position for serves, and sideways for ground strokes. For each stroke,
discuss how the extra weight improves performance (or not) in cases where the
player hits the ball off-center in (a) the left-right horizontal direction; (b) the
up-down vertical direction.
2.23 T Suppose the defender in Example 2.3 contacts the runner in the chest.
In terms of torque, explain why this is less effective than grabbing an ankle.
2.24 T To increase MOI, an object’s mass can be increased or the distance of
its mass from the center of rotation can be increased. Discuss the relative impact
of changing mass and changing distance in changing MOI.
2.25 T Golfers used to place metal tape around the edge of the backs of their
irons. Explain why this could improve performance on off-center hits. (Look up
pictures of old “muscleback” and newer “cavity back” irons to see how manu-
facturers responded.) Discuss where metal tape should be placed to improve the
performance of a golf putter.
 Rotational Motion 37

2.26 T One of the classification criteria for snowboard tricks is whether the
rider grabs the board or not. In terms of MOI, does grabbing the board make
any difference?
2.27 T In midair, an athlete swings her right arm across her body to her left
shoulder. By conservation of angular momentum, which direction will her body
rotate in response?
2.28 T Skater Brian Boitano had a signature jump (the “Tano triple”) in which
he raised his left arm straight into the air while executing a triple rotation. In
terms of MOI, does a raised arm make the jump easier or harder?
2.29 T In most of the situations we have discussed, larger MOIs are better.
Discuss situations in which a smaller MOI is better.
2.30 C Assume that a baseball bat has a constant mass density of c slug/in3 .
7
With the bat aligned along the x-axis, it has an outline given by y(x) = 192 x + 12
11
in for 0 < x < 24 and y(x) = 8 in for 24 < x < L for a bat of length L in.
(The barrel of the bat has constant thickness beyond 24 inches.) The cross-
sectional area at location x equals cπy(x)2 in2 , and the moment of inertia for a
RL
swing about the y-axis is 0 x2 cπy(x)2 dx. (a) Show that I ≈ cπ(0.63L3 − 2419).
(b) Show that at a constant √ angular acceleration α, the bat reaches θ = π
with angular velocity ω = 2πα rad/s. (c) From τ = Iα, conclude that α =
k/(0.63L3 −2419)
√ for some constant k. (d) Combine parts (b) and (c) to conclude
that ω = m/ 0.63L3 − 2419 rad/s for some constant m. (e) Watts and Bahill
measured rotation rates for major league batters and found that√ω ≈ 48 − .34L
rad/s. Compare the graphs of f (x) = 48−.34x and g(x) = 4000/ 0.63L3 − 2419
for 25 < x < 40 to see if the two calculations match well. (f) Which assumption(s)
in our calculation of ω are invalid?
2.31 C Calculate the MOI of a circular tennis racket frame that is 2 cm thick.
2.32 C Set up an integral for and then approximate the integral for the MOI
of an elliptical racket frame that is 1 cm thick and has (inner) dimensions of 11
cm by 16 cm.
2.33 P Find information about the degrees of difficulty for different dives. Do
the degrees of difficulty for different positions correspond to the MOIs for those
positions?
2.34 P A model of a baseball bat is given in exercise 2.30. Modify the model
to mimic a hollow aluminum bat. For wooden and aluminum bats of the same
dimensions and same weight, how do the MOIs for swinging compare? How do
the center of masses compare?
2.35 P Explore the usefulness of “corking” a baseball bat. This (illegal) ploy
involves hollowing out a cylindrical piece at the end of the bat and filling it with
cork (others have used bouncy balls). How much could the MOI change? How
much would this increase bat speed? Discuss whether the liveliness of the bat
(the coefficient of restitution discussed in Chapter 4) might decrease.
38 Sports Math

Further Reading
Peter Brancazio’s Sport Science is an excellent introduction to the physics
of sports.
Robert Watts and Terry Bahill’s Keep Your Eye on the Ball is the best of
several physics of baseball books.
A search for “Physics of (favorite sport)” will bring up many results.
I can recommend The Physics of Basketball (Fontanella), The Physics of
Hockey (Hache), Football Physics (Gay), Newton’s Football (St. John and
Ramirez), The Science of Soccer (Wesson), The Physics and Technology of
Tennis (Brody, Cross, and Lindsey), Golf Science (Smith), The Science of
Golf (Wesson), Golf By the Numbers (Minton), and The Physics of NASCAR
(Leslie-Pelecky).
The Howard Head story is told in Lee Torrey’s Stretching the Limits.
An impressive figure skating spin (titled “World Record Figure Skating
Spin”) can be found at https://www.youtube.com/watch?v=AQLtcEAG9v0
accessed July 15, 2015.
Further suggestions can be found in the notes at the Sports Math web site.
Chapter 3
Sports Illusions

Introduction
We have all experienced it. Trying to
kick or catch or hit a ball, we miss. And
then we hear the well-intended but conde-
scending voice of the coach: “Keep your
eye on the ball.” As we will see, in the
most literal of terms this advice cannot
be heeded. That is not to say that the ad-
vice is bad. Maintaining eye contact for as
long as possible is helpful, but it turns out
that we humans are not equipped with
enough visual acuity to track balls well.
“Keep your eye on the ball” is one of several mythbusting situations that
we explore in this chapter. We also try to answer the following questions. Are
softball pitchers better than baseball pitchers? Are great athletes born, and
not made? Should your football team take a penalty to improve the angle on a
field goal? Are referees fully objective and accurate judges? Do high jumpers
actually clear the bar? The myths in this chapter are all results of sports
illusions, shortcomings in the way our brains process images.

You Can’t Keep Your Eye on the Ball

A baseball pitcher throws a fastball across home plate. The batter, follow-
ing time-honored advice, tries to keep his eye on the ball. (Note that the exact
same situation occurs in tennis, softball, and other sports.) The ball, batter’s
eye, and the middle of home plate form a triangle that changes as the ball
moves toward home plate. Assuming that the batter stands 2 feet from the
plate, viewed from above the triangle looks like Figure 3.1.

39
40 Sports Math

The angle θ can be used to track the batter’s

gaze as he follows the ball. If θ changes rapidly,
then the batter’s eye must move rapidly to
x
keep pace. Using basic trigonometry, tan θ = ,
2
where x is the distance between the ball and the
plate. Calculus shows that the rate of change of
the angle θ (in calculus, we call it the derivative
of θ and denote it by θ0 ) is given by
2s
θ0 =
4 + x2
where s is the speed of the ball in ft/s. (Tech-
FIGURE 3.1: Eye on Ball
nically, the rate should be negative, since the
angle decreases. We are interested in the size of the eye movement, not its
direction, so we take the absolute value of rate.) We can now do a simple
calculation.

Example 3.1 For a 90 mph fastball, find the maximum value of θ0 , the rate
at which the batter’s eyes must move to watch the ball.
Solution. We first convert 90 mph to ft/s, so that the pitch speed is s = (90
mph) (5280 ft/mi)/(3600 s/hr) = 132 ft/s. Our calculus equation then gives
264
us θ0 = rad/s when the ball is at position x. What is the largest that
4 + x2
this can be? It’s clear that the smallest that x2 can be is 0, occurring at x = 0
(when the ball reaches home plate). So the smallest that the denominator can
be is 4. The smaller the denominator, the larger the number, so the maximum
264
value of θ0 is rad/s = 66 rad/s. This occurs at x = 0.
4

At this point, you may be underwhelmed, but hang on - there’s a great

punch line coming. Before getting there, let’s do a quick reality check on
Example 3.1. Does it make sense that the maximum rate of change occurs at
x = 0? Think about keeping your eye on a race car down the straightaway.
You have to move your eyes quickly to follow it as it nears you, but you will
have to move your eyes fastest when the car is even with you. The conclusion
in Example 3.1 makes sense.
The importance of Example 3.1 depends on the following physiological
fact: humans cannot accurately track objects that require angle changes at
rates of 3 rad/s or more. The 66 rad/s in Example 3.1 is completely out of the
question! In other words, keeping your eye on this ball is impossible.
(Actually, it might be possible to be looking in the right direction at all times;
even so, what you would “see” would be an out-of-focus blur far worse than
the ball in the picture at the beginning of the chapter.)
What do major league batters do about this? Watts and Bahill found two
strategies, both starting with a smooth tracking of the ball as far as possible.
 Sports Illusions 41

Some batters then continued moving their eyes at a constant rate, falling
behind the ball. Others moved their eyes forward to the place where they
predicted that the ball would cross the plate. There is some evidence that
excellent hitters are able to track pitches farther than poor hitters. So, “keep
your eye on the ball as long as possible” is good advice.
Players now have a variety of video games and training machines designed
to improve the players’ ability to track the ball and identify clues as to which
pitch is coming. The red seams on a baseball have characteristic appearances
for pitches, such as a red dot appearing on a slider and apparent train tracks for
two-seam fastballs and curveballs. As we will see, this type of visual training
can be critical.

You Can’t Touch This

The following simple calculation leads to an obvious conjecture. Surprising
theories of athletic excellence follow when we find that the conjecture is wrong.

Example 3.2 Ignoring air drag, compare the times for the following two
pitches: (a) a major league fastball at 95 mph from a distance of 60 ft; (b) a
women’s fastpitch softball pitch of 65 mph from 43 ft.
Solution. These are basic “time equals distance divided by rate” calculations,
but we do need to convert units. We have 95 mph (5280 ft/mi) ÷ (3600 s/hr)
≈ 139.3 ft/s for (a) and 65 mph (5280 ft/mi) ÷ (3600 s/hr) ≈ 95.3 ft/s for
(b). The times are then (a) 60 ft ÷ (139.3 ft/s) ≈ 0.43 s and (b) 43 ft ÷ (95.3
ft/s) ≈ 0.45 s.

Arguments can be made about

whether a 95 mph fastball is rarer
than a 65 mph softball pitch, and we
have seen that air drag changes time of
flight significantly. However, the point
is that the amounts of time available
for batters in major league baseball
and women’s softball are nearly equal.
It would then be an obvious conclusion
that major league batters should have
no trouble hitting softball pitchers (es-
pecially since softballs are larger than
baseballs). David Epstein’s book The
Sports Gene describes what happened when this was put to the test. In 2004
and 2005, USA Olympic pitcher Jennie Finch recorded a series of batting
practice encounters with baseball’s best hitters. Mike Piazza was blown away;
42 Sports Math

Albert Pujols could not get his bat on a ball. Barry Bonds was almost too
befuddled to swing. Alex Rodriguez watched five warm-up pitches and left
with the remark, “No one’s going to make a fool out of me!” (Which turned
out to be not very accurate.)
If the reaction times are the same, then why could the players not hit
Finch? An essential part of being a successful batter is to increase the reac-
tion time beyond the 0.4 seconds or so of a pitch. As Finch proved, that is
not enough time to see and react. Instead, over time batters develop an un-
conscious catalog of clues that enable them to predict the speed and location
of pitches before they leave the pitcher’s hand. Given no catalog of clues to
Finch’s delivery, even the best and quickest hitters were hopeless, to the point
of almost never even getting a bat on the ball.
Some clever experiments have shown that these clues exist. If you are
shown a video of a pitch that stops right before the ball is released, could you
predict the speed and location of the pitch? The better the batters, the more
accurate their predictions are. Can we determine which clues are being used?
Various occlusion experiments doctor the video to black out, for instance, the
pitcher’s shoulders. If the accuracy of your predictions decreases, then there
is evidence that you are using something about the motion of the shoulders.

You Can’t Teach Size

The preceding discussion indicates that the experience and practice of
building the catalog of clues is more important than the raw speed or visual
acuity of the athlete. Questions of how much of player success is genetic and
how much is due to environment (including hard work) are intriguing but hard
to answer.
An important addition to the Jennie Finch study, also reported in The
Sports Gene, is that in fact the top baseball players and softball players have
outstanding eyesight. Actually, “outstanding” is an understatement, as major
leaguers and 2008 U.S. Olympians tested at an average visual acuity of 20/11,
with superior depth perception and contrast sensitivity to boot. Of the Los
Angeles Dodgers tested, about 2% scored a 20/9 acuity rating, considered
to be near the limit of human capacity. The top hitters are equipped with
outstanding genetics as well as well-developed brains.
There are no simple answers. Almost nobody succeeds in sports without
long hours of dedicated practice (or, if you’re Allen Iverson, game action).
You can’t hit Jennie Finch without many reps against her or similar softball
pitchers. An oversimplification that appears in many sound bites is the 10,000
Hour Rule. This says that top performers in almost any field need to have
logged at least 10,000 hours of deliberate practice before reaching elite status.
The term “deliberate” encompasses much. Your practice can’t be dull and
 Sports Illusions 43

plodding, just putting in the hours; it must be focused practice that actively
addresses any weaknesses that you may have. The easiest way to characterize
deliberate practice is to say that this is what someone completely dedicated
to becoming the best of all time would do. Unfortunately, this makes our
definition circular.
Contrary to a common misinterpretation of the 10,000 Hour Rule, putting
in the hours does not guarantee success. When I was a college student, putting
in 10,000 hours at a basketball camp teaching low-post skills would not have
made me an NBA-caliber player. My (mostly) genetic endowment of 5’11”
height and no jumping ability precluded that. However, all of that practice
would have helped me abuse the other students in the pick-up games in the
gym.
An interesting question in the sports version of “nature versus nurture” is
whether hard work, persistence, or love of the game have a genetic component.
Sons of football players often love football, but that could be easily explained
by an environment of positive experiences with football. A related issue in-
volves the effect of training. Some people respond dramatically to training,
increasing strength or speed in leaps and bounds. For others, the same training
produces no improvement. Is “trainability” genetic? The rapid conversion of
Usain Bolt from a high school cricket and football player to the world’s fastest
sprinter ever seems to defy the 10,000 Hour Rule. Bolt responded to standard
training techniques with spectacular improvements in time. Is his greatest
gift a speed gene or a trainability gene? Either way, his accomplishments are
remarkable.

You Can’t Afford the Yardage

A football player is driven out of bounds
just two yards from the goal line, bringing up
fourth down. The coach calls for a field goal,
after taking a 5-yard penalty to “improve the
angle.” What is wrong with this picture? You
might first object to the conservative choice of
trying a field goal; we explore the many pros
and few cons of going for it in Chapter 9. From
the standard television camera behind the goal-
posts, the angle for a short field goal from a
hash mark does look severe. But, does it really FIGURE 3.2: Field Goal
increase the angle to move back 5 yards?
44 Sports Math

Example 3.3 The goal posts in college football are 18 ft 6 in apart and the
hash marks are 40 ft apart. Compare the angle for a field goal from the right
hash mark from distances of (a) 20 yards and (b) 25 yards.
Solution. With the ball on the right hash mark, draw a triangle with vertices
at the ball, the place where the hash mark would hit the back of the end zone,
and the far goal post. (See Figure 3.2.) This is a right triangle with sides of
20 yards (in part (a)), 29 ft 3 in (20 ft from the hash mark to the center of
the goal post plus 9 ft 3 in more), and the length of the hypotenuse (which we
will not use). In a common unit of feet, the sides are 60 and
29.25. The angle
−1 29.25
θ at the ball satisfies tan θ = 29.25
60 and so θ = tan 60 ≈ 0.4536 rad =
0.4536 rad × 180/π deg/rad ≈ 25.99 deg. To compute the angle to the near
goal post, we only change the distance from the hash mark to
the near post to
20 − 9.25 = 10.75 ft. The angle changes to θ = tan−1 10.75 60 ≈ 0.1773 rad =
0.1773 rad × 180/π deg/rad ≈ 10.16 deg. The margin of error for the kick, the
angle between the two posts, is approximately 25.99 − 10.16 deg = 15.83 deg.
For part (b), we change the two calculations in part (a) by
replacing 60 with
75. The angle between the two posts is now tan−1 29.25 −1 10.75


75 − tan 75 ≈
0.2295 rad ≈ 13.15 deg, 2.7 degrees less than the shorter kick.

Moving 5 yards from a 20-yard field goal to a 25-yard field goal reduces the
angle for a successful kick by nearly 17%. Why have teams thought for years
that this was good strategy? An argument can be made that the angle is not
the only variable to take into account. From the middle of the field, the kick
is perpendicular to the line of scrimmage and directly into the crowd behind
the goal post; numerous visual cues help you line up the kick. From the hash
mark, these cues are largely gone, and lining up may be more difficult. If there
is a decrease in accuracy due to the odd look of the short kick, perhaps that
makes up for the loss in accuracy due to a decreased angle of success.
A different situation sometimes occurs at the end of football games. A
team running the clock down before trying a game-winning field goal may run
one last play where the quarterback takes the snap, slides over to the center
of the field, and kneels. The play may lose a yard or two, but getting the ball
into the center of the field is considered worth it. From the perspective of
angles, is this true?

Example 3.4 Compare the angles for a college football (a) 35-yard field goal
from the right hash mark and (b) 37-yard field goal from the center.
Solution. For (a), we repeat the analysis from Example

3.3 with 60 ft replaced
by 105 ft. The angle is tan−1 29.25 −1 10.75


105 − tan 105 ≈ 9.72 deg.
For (b), we must start over. Consider the two triangles to be drawn, one
to the goal post on the left and the other to the goal post on the right. In each
case, the leg of the triangle along the goal line has length 9.25 ft. Thus, the
angles to the two

posts are the same; the angle of success is twice this angle,
or 2 tan−1 9.25
111 ≈ 9.53 deg. In this case, dropping back two yards decreases
the angle more than centering the ball increases the angle.
 Sports Illusions 45

You will discover in exercise 3.8 how much centering the ball improves the
angle, and how far back the quarterback can kneel before the overall angle is
decreased.
It should be noted, and will be explored in the exercises, that college
goalposts and hashmarks are wider than those in the NFL, and narrower than
those in high school. The numerical conclusions drawn will be different with
different field specifications, and the interpretations vary with the differences
in abilities of the kickers.

You Can’t Bend That Way

In Chapter 2, we talked about the ad-
vantages of a long swing in golf or tennis. I
once played against a golfer who essentially
had no backswing; he raised his club into
a good position and then started his down-
swing. Why do golfers take vigorous back-
swings, and pitchers use complicated wind-
ups? The picture to the right gives a clue.
In Figure 3.3, the clubhead has stopped
going back but has not yet started forward.
The clubhead’s velocity, then, is zero. So why
not use my opponent’s strategy and simply
place the club in that position? The answer
is that you could place the clubhead in that FIGURE 3.3: Club Bend
location, but you would not get the bending
of the club. The bending/unbending of the club works enough like a spring
that it is worth a quick detour to discuss energy in a spring.

Example 3.5 Suppose a spring is compressed by 2 m and then released.

Track its energy as (a) it moves through its resting position and (b) reaches
its maximum stretching position.
Solution. At the beginning of this process, the spring has zero kinetic energy
(K = 21 mv 2 ). However, after its release it picks up speed and therefore kinetic
energy. In its initial position, we say that the spring has potential energy U to
indicate that it has the potential for an increase in kinetic energy. (a) A spring
at its resting position has zero potential energy; if it starts with no kinetic
energy, it will remain stationary. However, in our situation it has accumulated
kinetic energy. So, U = 0 and K > 0. At position (b), the spring has stopped
stretching and not yet started to compress; therefore, it has no kinetic energy.
However, it now has potential energy (the spring will compress and regain
kinetic energy). In the absence of friction, conservation of energy tells us that
46 Sports Math

the sum K + U will remain constant, so that the potential energies at the
beginning and at position (b) will be the same, with both equal to the kinetic
energy at position (a).

In Figure 3.3, the golf club has little kinetic energy, but the bending of the
club gives it potential energy; like a compressed spring, it will snap back to
its resting position, creating kinetic energy. This potential energy translates
into extra clubhead speed, and is created by the transition from backswing to
downswing. This is an important advantage of an energetic backswing.
Similarly, a pitcher’s elbow undergoes unnatural stresses as the pitcher
transitions from arm cock to follow through. The following example gives an
idea of how much force is applied.

Example 3.6 Research indicates that up to 120 N-m of valgus torque (con-
sidered the prime cause of elbow injuries in pitchers) is exerted on a major
league pitcher. At a distance of 13 m (about the distance from hand to elbow),
how much force is needed to create this torque?
Solution. Recalling that τ = F L, we want 120 N-m = (F )( 31 m) so that
F = 120 ∗ 3 = 360 N. For comparison purposes, 360 N is about 81 pounds,
enough force to create significant potential energy in the arm to increase
throwing speed.

You Can’t Make That Call!

Sports referees have difficult and typically thankless jobs. One of the hard-
est referee decisions in sports is the offside/onside call in soccer/football. The
referee must keep multiple players and positions in mind to make the correct
call. Research showed that in the 2002 World Cup mistakes were made on
offside calls an astonishing 26% of the time. Improved research and training
reduced that percentage to under 10% in the 2006 World Cup.
In the most common scenario, a player
passes the ball ahead to a teammate (an at-
tacker). At the instant the ball is passed, imag-
ine stopping the action and drawing imaginary
lines through each player and parallel to the
goal line. Typically, the goalkeeper’s line will
be closest to the goal line. The next closest line
through a defender (called the last defender) is
the offside line. If the attacker’s line is closer FIGURE 3.4: Offside
to the goal line than the offside line, then the
attacker is offside, the play is stopped, and the
 Sports Illusions 47

defending team is awarded a free kick. In Figure 3.4, an attacker at position

A is onside, while an attacker at position A’ is offside. The AR (assistant ref-
eree, or sideline referee) is tasked with monitoring which defender is the last
defender and whether attackers are onside or offside.
There are variations on the rule, but we will work with this basic situation
and examine three possible ways for the AR to unknowingly make an incorrect
call.
The first possibility to consider is the shift-of-gaze error. If the AR is watch-
ing the ball until it is kicked, and then locates the attacker and last defender,
during the time required for the AR to re-focus the attacker could move from
position A (onside) to position A’ (offside). Attackers try to time their run so
that they pass the last defender shortly after the ball is kicked, giving them a
head start on running to the pass. As seen in Example 3.7, the referee could
see the attacker as being offside when, at the exact instant of the pass, the
attacker was onside. However, research shows that elite referees always watch
the attackers and defenders, tracking the ball only through peripheral vision.
At the highest level, shift-of-gaze errors rarely occur.
A second possibility is optical error. Re-
search has shown that ARs are only rarely po-
sitioned on the offside line. They are typically a
meter or two closer to the goal than the offside
line. Because the AR’s line of sight is not paral-
lel to the offside line, mistakes such as the one
in Example 3.7 can be made. In studies, slightly
FIGURE 3.5: Optical
less than half of the offside mistakes are consis-
tent with optical error. A higher percentage of
errors can be explained by the flash-lag effect,
to be discussed after a quick example.

Example 3.7 Show that incorrect calls are made in each of the following
situations. (a) (Shift of Gaze) The attacker in Figure 3.4 is onside at position
A, one-half meter behind the offside line and running at a pace of 10 m/s; the
defenders are not moving. If it takes the AR 0.1 seconds to shift gaze from
the ball being passed to the attacker, will the AR make the correct call? (b)
(Optical) The attacker in Figure 3.5 is offside at position B, 0.3 m in front of
the offside line and 30 m from the touchline (sideline). The last defender is 40
m from the touchline, and the AR is standing 2 m in front of the offside line.
Will the AR make the correct call?
Solution. (a) At a speed of 10 m/s, the attacker will advance 10(0.1) m = 1
m in 0.1 seconds. The attacker has moved from 0.5 m behind the offside line
to 0.5 m in front of the offside line, and will be called offside.
(b) Think of a coordinate system where the defender is the origin and
the offside line is the x-axis. The line from the defender to the AR has slope
2
40 = 0.05 and has equation y = 0.05x. This is the AR’s effective offside line;
everybody below this line will be judged onside. The attacker is 10 m closer to
48 Sports Math

the touchline, and thus is at x = 10. At x = 10, the offside line passes through
y = 0.05(10) = 0.5. The attacker at 0.3 m is below this line, and thus appears
onside to the AR.

We have seen two ways for referees to miss this difficult call. A physiolog-
ical phenomenon called the flash-lag effect presents another challenge. In the
television series Brain Games, this effect is presented as a game in which a
football (American style) moves at constant speed across the screen. At some
time, a red dot flashes in the upper-right corner; we viewers are supposed to
determine the position of the football when the flash occurs. The interesting
result is that 95% of test subjects misplace the football, believing that it has
gone farther across the screen than it has. The explanation is that it takes
about one-tenth of a second for the brain to process the surprise flash, and
the brain marks the location of the football after that processing occurs. As in
Example 3.7, during this time lag the football moves to a new location, which
we incorrectly “see” as the location at the time of the flash.
Substitute moving football players for a moving football and a kick for a
red dot and you have the main components of the offside call. Interestingly,
research has shown that good referees make mistakes that are consistent with
flash-lag effect, but elite referees often do not. They have learned some way to
correctly compensate for the flaw in their human vision processing systems.
An interesting variation of this result oc-
curs in baseball. The play in question is at first
base, when a throw from an infielder pops into
the first baseman’s mitt at the same time as
the runner’s foot hits the bag. Is the runner
safe or out? I’ve always been amused at the
coolness of the umpires, who will look at the
play and eventually raise a fist casually to in-
dicate that the runner is out (in fact, clearly
FIGURE 3.6: Out?
out, don’t even think about arguing).
Research shows that umpires raise their fists too often, calling runners out
who are actually safe at first. The reason for this is rooted in how the human
brain processes sight and sound. Major league umpire Mike Winters describes
an umpire teaching school technique of blindfolding umpires. This is not to
validate the loud fan in the thirtieth row who thinks a blind person could
make more accurate calls. This is to teach umpires to listen for the pop of the
ball in the mitt. Winters says that umpires who listen for the ball pop and the
thud of the runner’s foot on the bag never miss the call, but those who try to
visually determine whether the ball reaches the mitt before the foot touches
the bag often make mistakes.
Umpires are taught to keep their eyes on the base while listening for the
ball to hit the mitt. This way they avoid the shift-of-gaze and flash-lag er-
rors discussed above. However, this runs the umpire straight into a different
cognitive illusion. Over 100 years ago, German experimenter Wilhelm Wundt
 Sports Illusions 49

discovered that when an audio event is paired with a visual event, the brain
system that syncs the two tends to pre-date the visual event. Thus, if the
ball pops into the glove at the same time as the foot hits the bag, the brain
backs the runner up a small amount and syncs that previous position with
the sound. The umpire perceives the sound as occurring before the runner hits
the bag, and makes the out call.
There is a clear advantage for the brain to work this way. Think about a
door being slammed from a long distance away. The sight of the door closing
reaches your eyes before the sound of the door shutting reaches your ears (light
travels faster than sound). An intelligent brain backtracks the visual image to
correctly sync sight and sound. Unfortunately, this brain feature becomes a
bug when trying to make the correct call at first base.
On June 2, 2010, Detroit pitcher Armando Galarraga found out that the
Wundt effect is not always active. Jason Donald hit a ground ball to Detroit
first baseman Miguel Cabrera, who flipped the ball to Galarraga, beating
Donald to the base for the out ... except that umpire Jim Joyce called the
runner safe. The tragedy is that an out would have been the 27th and final
out of a perfect game.
Think about Joyce’s situation: with the Detroit crowd screaming for a
perfect game, he had to hear a soft flip from Cabrera hitting Galarraga’s
glove. In fact, Galarraga snagged the ball in the webbing of his glove, so there
was likely no noise at all. If Joyce, known as an excellent umpire, had to
rely on eyesight only, the call was made difficult by all of the factors that
afflict assistant referees in soccer. The bottom line was a very bad bit of luck
for Galarraga, although the extra publicity for the bad luck and widespread
praise of the great sportsmanship shown by both Galarraga and Joyce may
have made up for the loss of a perfect game.

You Can’t Clear That Bar

The final example in this Chapter (more to follow in the exercises, though!)
involves high jumping. If you are old enough, you may remember the straddle
technique that was routinely used until the 1970s. If not, search for an old
video; the technique looks very odd to modern eyes. High jumpers now use
a technique introduced to the world by Dick Fosbury in the 1968 Olympics
(and therefore known as the Fosbury Flop) where they jump backward, arch
their backs and then pull their legs over the bar. A quick look at the winning
jumps in various Olympics (Figure 3.7) suggests that success with the straddle
technique had plateaued by 1976, and the flop technique took it to a new level,
at which we are now stuck. (To be fair, a plot of world record jumps tells a
different story.)
50 Sports Math

FIGURE 3.7: Heights of Winning Men’s High Jumps, Olympics, 1960-2012

Here’s the amazing part: in terms of the

Newtonian mechanics outlined in Chapter
1, the best high jumpers do not clear the
bar! To be precise, the center of mass of the
high jumpers can pass underneath the bar
on a successful jump (recall that our equa-
tions in Chapter 1 only track the center
of mass). How can this be? Take a careful
look at the jumper in Figure 3.8. Her head
has cleared the bar and is now underneath
FIGURE 3.8: High Jump
the height of the bar, as are her shoulders,
lower legs and feet. More of her mass is below the bar than above, and her
center of mass (where the sums of mass times distance for her body parts
balance) is below the bar. She has performed a magic trick, clearing a bar
that is higher than she can jump.

Exercises
In these exercises, T refers to thinking problems, conceptual problems re-
quiring no calculations. C refers to problems requiring calculus or significant
computer calculations. P refers to projects; these are ideas for further inves-
tigation (hints and resources are at the book’s web site).

3.1 Rework Example 3.1 with a 75 mph curve ball.

3.2 In Example 3.1, (a) find the maximum ball speed at which a human could
track the pitch all the way to home plate; (b) determine how far the batter can
track the pitch (i.e., the smallest x such that the rotation rate is 3 rad/s or less).
 Sports Illusions 51

3.3 A tennis player tries to track a serve that is on a line 2 feet to the side at the
speed 105 mph. (a) What is the rotation rate needed to track this serve all the
way to the racket? (b) Given a maximum tracking rate of 3 rad/s, how far can
the player track the serve? (c) Can a linesperson sitting 30 feet to the side track
the ball?
3.4 A spectator sits 300 feet away from a racetrack and tries to follow a race car
moving at 180 mph. (a) What is the rotation rate needed to track the car all the
way? (b) Given a maximum tracking rate of 3 rad/s, what is the closest the fan
can sit to the track and completely track the car?
3.5 Ignoring air drag, find the reaction times for the following: (a) a tennis serve
at 130 mph from 78 feet away; (b) a lacrosse shot at 100 mph from 30 feet away;
(c) a penalty kick at 80 mph from 12 yards away; (d) a hockey shot at 90 mph
from 30 feet away; (e) a line drive at 150 mph at a third baseman 85 feet away.
3.6 Data from MLSsoccer.com indicates the following breakdown of birth months
for United States U-17 and U-20 national team players: 14.1% (Jan), 10.8%
(Feb), 12.0% (Mar), 9.9%, 8.5%, 7.0%, 6.0%, 7.7%, 6.0%, 7.1%, 4.5% (Nov),
6.5% (Dec). If all birth months were equally likely, what percentage would each
month have? What is the average absolute difference between expected and
actual percentages? “U17” means players must be under the age of 17 as of
January 1. Compare the ages of the oldest U17 player and the youngest (of
those who would not qualify for U16). Does birthdate matter?
3.7 Repeat Examples 3.3 and 3.4 for (a) high school dimensions of goal posts 23
ft, 4 in apart and hash marks 53 ft, 4 in apart; (b) NFL dimensions of goal posts
18 ft, 6 in apart and hash marks 18 ft, 6 in apart.
3.8 In Example 3.4, (a) how many yards lost could a team absorb and still have
an equal angle for kicking? (b) Repeat for high school kickers.
3.9 At the top of a pole vault, a vaulter of mass m kg at height h m has grav-
itational potential energy of mgh J. The kinetic energy of the vaulter before
planting the pole is 12 mv 2 where the vaulter’s running speed is v m/s. Assuming
that these are equal, and that the top running speed of a vaulter is 12 m/s, find
a quick estimate of the maximum height of a pole vault. The world record in
2014 is 6.16 m; how does your estimate compare?
3.10 In Example 3.7 (a), if the attacker is onside by 1.5 m and the defender is
also running at 10 m/s but in the opposite direction, show that an AR with a
0.1 s gaze shift will get the call wrong.
3.11 In Example 3.7 (b), determine the set of positions (distance on- or offside
and distance from the touchline) in which (a) the AR incorrectly sees an offside
attacker as being onside; (b) the AR incorrectly sees an onside attacker as being
offside.
3.12 Repeat exercise 3.11 if the AR is only 1 m ahead of the offside line.
3.13 The definition of offside involves the location of the attacker at the time
of contact with the ball. If this contact lasts 0.05 s, determine circumstances in
which the attacker could be legally both onside and offside.
3.14 Plot a graph similar to Figure 3.7 but using world record high jumps instead
of Olympic winning heights. In your new graph, does 1968 appear to be an
important year? Which graph, yours or Figure 3.7, do you think more accurately
reflects the evolution of the high jump? Give reasons.
52 Sports Math

3.15 T Baseball batters talk about “rising fastballs” and curve balls “dropping
off the edge of a table.” For a batter who redirects his eyes to where he predicts
the pitch will end up, explain how this illusion could occur.
3.16 T Ted Williams, considered by many to be the greatest hitter ever, claimed
that he could see the bat hit the ball. His eyesight, measured in standard terms
to be 20/9, was extraordinary but do you think he could track the ball all the
way to home plate? If not, would it be possible for him to clearly see the contact
of bat and ball?
3.17 T Submarine-style pitcher Chad Bradford threw underhanded, sometimes
scraping his fingers on the ground. Even though Bradford did not throw hard
(85 mph or so), explain why batters might have trouble hitting his pitches.
3.18 T Michael Jordan famously took time off in the middle of his NBA career
to play professional baseball. Even though Jordan was a fabulous athlete, explain
why it would be reasonable to expect him to struggle as a hitter.
3.19 T Hold a ruler or meter-stick vertically, have a friend place two fingers on
either side of the stick, drop the stick, and have your friend grab it as quickly as
possible. Explain how this can be used to determine your friend’s reaction time.
3.20 T The television show Sport Science tested drag racer Hillary Will’s re-
action time (critical to racers) using the starting sequence of descending lights
used in races. She was clocked with a reaction time of 0.001 s. Human reaction
times for seeing a green light and pushing a button are more typically on the
order of 0.2 s. Explain this apparently superhuman performance.
3.21 T You are familiar with the “on your mark, get set” BANG that starts
running races. To try to guarantee that sprinters are reacting to the gun and not
anticipating the start, there are harsh penalties for false starts and a “reaction
time” of less than 0.1 s is considered a false start. Discuss how reaction time
might be measured and why 0.1 is the threshold.
3.22 T Mo Farah of Great Britain won the 10,000 m race at the 2012 Olympics.
The previous four Olympics had been won by Ethiopians. For an event dominated
by a country or region, give reasons why the domination might be mostly nature
(genetics) and reasons why the domination might be mostly nurture (training
and motivation).
3.23 T The armspan-to-height ratio for most humans is about 1 (measure your-
self!). The average ratio in the NBA is 1.063; while this may not seem much dif-
ferent than 1, a ratio higher than 1.05 triggers tests for Marfan’s disease. Discuss
the advantages for an NBA player to have a high ratio.
3.24 T The human body shows remarkable abilities to adapt to training and
improve at specific tasks. We sometimes underestimate the specificity: for exam-
ple, improvements in lifting free weights do not always transfer to other measures
of strength. Discuss the benefits of swinging a heavy bat in warmup or wearing a
high-drag suit for swimming training. Explain why these do not necessarily help
batting and swimming performance.
3.25 T Discuss ways in which a tennis serve motion could bend the server’s
elbow unnaturally, and why this could increase serving speed.
3.26 T If an attacker is running toward the goal, under the flash-lag hypothesis
discuss in which situation the AR is more likely to miss an offside call: (a) the
 Sports Illusions 53

defender is stationary; (b) the defender is running away from the goal in an
offside trap.
3.27 T Test yourself on the Brain Games football challenge: season 2, episode
11, Illusion Confusion.
3.28 T Explain why the tendency of baseball umpires to call runners out at
first base when they are actually safe is contrary to the flash-lag theory.
3.29 T In Figure 3.6, in 1980 there is a significant increase in winning height.
Give a reason that the introduction of the Fosbury Flop in 1968 might explain
this increase.
3.30 C Given that tanθ(t) = 21 x(t) and x0 (t) = s, derive the rate of change
2s
equation θ0 (t) = .
4 + x2
3.31 C Generalize exercise 3.30 to the case where tanθ(t) = L1 x(t) for some
length L.
3.32 C Adapt the baseball equations of motion (x0 = vx , y 0 = vy , z 0 = vz ,
vx0 = −fd vx vx2 + vy2 + vz2 + fm (wy vz − wz vy ), vy0 = −fd vy vx2 + vy2 + vz2 +
p p

fm (wz vx − wx vz ), vz0 = −32 − fd vz vx2 + vy2 + vz2 + fm (wx vy − wy vx ), fd =

p

0.002203, fm = 0.000632) to model a pitch that starts at the point (55, 5) and
heads toward home plate at (0, 0). Assuming pure backspin and an initial velocity
that is horizontal at 95 mph, find the spin rate in rpm that would actually
produce a rising fastball. (Hint: you need the initial vz0 to be positive.)
3.33 C Referring to exercise 3.32, simulate the paths of fastballs that are thrown
horizontally (a) at 95 with 1500 rpm backspin; (b) at 100 mph with 1600 rpm
backspin. (c) Find the height of each pitch when it is 20 feet from home plate.
Would the batter be able to tell the difference? (d) Find the height of each pitch
when it reaches home plate. (e) If the batter uses the height 20 feet away to
predict the height at home plate, why might he think that the 100 mph/1600
rpm fastball hopped at the end?
3.34 C For the kicker in Example 3.3, find the distance x at which the angle is
maximized. Why is this not a valid distance in football?
3.35 C For kickers in high school (goal posts 23’4” wide and hash marks 53’4”
wide), find the distance x at which the angle is maximized. Is this a valid distance
in football?
3.36 C A hockey player races down the ice in a straight line parallel to the
sideboards and 3 feet wide of the net. The net is 6 feet wide. At what point is
his left-right margin of error at its maximum?
3.37 C A soccer player shoots from 20 yards out. Where should the keeper
be positioned to minimize the angle for scoring a goal? The goal is 24 ft wide;
assume that the keeper can cover 12 ft. Minimize the error if the shooter is (a)
4 ft wide of the left post; (b) 4 ft inside the post; (c) generalize your answers.
3.38 C Generalize exercise 36 to a shooter who is a ft wide of the net, with a
p
net that is b ft wide. Show that the maximum margin of error is at (a + b)b ft.
3.39 P Explore the advantages or disadvantages of being lefthanded in sports.
Possible research avenues are studies showing that (1) lefthanded athletes die
younger, or (2) lefthanded batters have higher batting averages, or (3) there are
a disproportionate number of lefthanders in elite tennis and baseball.
54 Sports Math

3.40 P Investigate one of the following claims: (1) football teams from warm-
weather cities do not perform well in the cold; (2) championships are won by
teams with strong defenses (not strong offenses); (3) at most universities, college
football brings in enough money to pay for non-revenue sports.

Further Reading
Robert Watts and Terry Bahill’s Keep Your Eye on the Ball is the source
for the idea that batters can’t keep their eyes on the ball, and for the research
of what batters actually do.
David Epstein’s The Sports Gene is an excellent read on many aspects of
the nature versus nurture debate.
Malcolm Gladwell’s Outliers gives an enjoyable overview of research into
elite performances in a variety of fields. K. Anders Ericsson is a proponent
of the 10,000 hour rule in sports; see “The Role of Deliberate Practice in the
Acquisition of Expert Performance” in Psychological Review (1993).
The Sloan Sports Analytics Conference generously posts videos of many
of its sessions. A discussion between Epstein and Gladwell on nature versus
nurture occurred in 2014, and is titled “10,000 Hours vs. The Sports Gene.”
Harold Klawans’s Why Michael Couldn’t Hit gives a wealth of information
about the neurological basis of sports performances.
More about umpires making the call at first base can be found
at http://www.nytimes.com/2011/06/11/sports/baseball/first-base-umpires-
call-them-as-they-hear-them.html accessed 7/23/2015.
Gordon Russell’s Sport Science Secrets discusses a variety of sports facts
and myths, including the close calls at first base and the home field advantage.
Sources for the offside call include “Offside decision making in the 2002 and
2006 FIFA World Cups” by Catteeuw, et al; “The Effects of Additional Lines
on a Football Field on Assistant Referees’ Positioning and Offside Judgments”
by Barte and Oudejans; “Visual Scan Patterns and Decision-Making Skills of
Expert Assistant Referees in Offside Situations” by Catteeuw, et al.
“Seeing the Benefit: MLB teams focus on enhancing players’ visual train-
ing” by Stephanie Apstein discusses players’ visual acuity and new training
techniques. Sports Illustrated April 13, 2015 issue.
Sport Science “Myths” episode with Jason Zuback, https://www.youtube
.com/watch?v=Ito3BSO-St8 ; “Long Driver” with Jamie Sadlowski,
https : //www.youtube.com/watch?v = oU ZeBzkcLU 0 accessed 7-24-2015.
Swimming training articles include http : //www.swimmingscience.net/
2013/08/drag − suits − part − ii.html, accessed 7-24-2015.
The Sport Science episode “Lose the Weight” explores the use of a weighted
bat, https : //www.youtube.com/watch?v = 0v R8UK rhY accessed 7-24-2015.
Chapter 4
Collisions

Introduction
Football is a sport of collisions. Fans and
players celebrate big hits as enthusiastically
as touchdowns. The same is true of hockey,
where a strong check always earns a roar of
approval from the fans.
The same is true of tennis, .... Did I
get your attention? Even a McEnroe-esque
screaming tantrum on the tennis court is un-
likely to have physical contact, so why would
I claim that tennis is a sport of collisions?
Here’s why: every point starts with a colli-
sion between racket and ball, followed by a
collision between ball and ground, and so on.
The motion of the ball, and therefore the outcome of each point, is determined
by these collisions.
In this chapter, we look at collisions of
many types. Some are the person-to-person
collisions that energize sports fans, but most
are collisions between the tools of the sports,
the commonplace occurrences that give the
sports their distinctive characters. We will
see that hockey, in addition to bone-crushing
checks, has the most intricate of collisions in
the slap shot’s dance of the stick and puck.
We will find answers to the following ques-
tions. How much force does a football player
absorb? Why does a kangaroo jump better
than a human? Can a baseball player hit a
ball farther by gripping the bat more tightly?
Where is the sweet spot on a tennis racket? How do sports regulate the equip-
ment to keep the games competitive? Why did Rick Barry shoot free throws
underhanded, and is it a coincidence that he is one of the best of all time?

55
56 Sports Math

Linear Momentum
In a 2014 college football game, Clemson defender Jayron Kearse tackled
Louisville running back Dominique Brown head-on (let’s say “shoulder-on” to
keep it legal), stopping him in his tracks. The play was unremarkable, except
for the fact that Brown weighed in at 230 pounds compared to Kearse’s 200
pounds. How is the smaller man able to win the collision? You probably know
that the answer has to do with speed, that the impact of a hit depends on
both the mass and the speed of the hitter. The follow-up question is to ask for
the exact combination of mass and speed (add them? multiply them? square
them?) to determine the outcome of a collision.
The answer, as usual, comes from Newton’s Second Law. Rather than the
form F = ma, replace acceleration with change in velocity (∆v) divided by
change in time (∆t). Multiplying across by ∆t, we have that

F ∆t = m∆v = ∆(mv)

where we assume that the mass m is constant. We give names to both sides
of this revised Second Law. The combination F ∆t is called impulse (J) and
the combination mv is called linear momentum (p). Our equation is then
the impulse-momentum equation

J = ∆p

or, in words, impulse equals the change in linear momentum.

When there is no force, the impulse is zero, making the change in linear
momentum zero. This is conservation of linear momentum:

In the absence of external forces, linear momentum is conserved.

This enables us to solve our football problem. Assume that just before the
collision Kearse is moving in one direction, which we will call the positive
direction, with speed vk . His linear momentum is +mk vk . Brown moves in a
negative direction with linear momentum −mb vb . The total linear momentum
of the two players before the collision is +mk vk − mb vb . Assume that there are
no external forces (no other players joining the pile, no pushing on the ground
by either player, and so on). Then the total linear momentum is conserved,
so that the total linear momentum before the collision equals the total linear
momentum after the collision. The situation after the collision is a single
Kearse/Brown tangle of mass mk + mb moving with a combined velocity of
vc . Then
mk vk − mb vb = (mk + mb )vc
where Kearse wins (the players move in the Clemson direction) if vc > 0 and
Brown wins if vc < 0.
 Collisions 57

We are almost there. Kearse wins if the total linear momentum is positive.
Using the total linear momentum before the collision, this means that mk vk −
mb vb > 0, which happens if mk vk > mb vb . So, we have a simple conclusion:
whichever player has the larger linear momentum wins!

Example 4.1 A 230-pound running back moving at 20 ft/s collides with a

200-pound defensive back. Assuming no external forces, determine how fast
the defensive back needs to be moving to stop the running back.
Solution. Since an object with mass m has weight mg for the gravitational
constant g, the masses of the two players are 230/g slugs and 200/g slugs,
respectively. Assume that the defensive back runs with speed v ft/s. By the
discussion above, the assumption of conservation of linear momentum (no
external forces) implies that the player with the larger linear momentum wins.
The defensive back stops the runner if v ∗ 200/g ≥ 20 ∗ 230/g. Multiplying by
g removes that constant, and we conclude that v ≥ 20 ∗ 230/200 = 23 ft/s.
The running back weighs 15% more, so the defender needs to be moving 15%
faster to stop him.

Impulse and Force

We know how fast Kearse needs to move to stop Brown, but how much force
does he deliver? The impulse-momentum equation, which states that impulse
equals change in linear momentum, is a place to start. This time, we focus
on just the running back Brown, for whom conservation of linear momentum
does not apply (since the defender is intent on delivering an external force to
him). He enters the collision with linear momentum mv and exits the collision
with linear momentum 0 (he has 0 velocity). Therefore, his change in linear
momentum equals mv − 0. Using the values from the example above, m =
230/g slugs and v = 20 ft/s, we have ∆(mv) = 4600/g slug-ft/s. The awkward
unit of slug-ft/s is equivalent to lb-s, which is not coincidentally a unit for
the impulse J = F ∆t. By the impulse-momentum equation, we now have
F ∆t = 4600/g ≈ 144 lb-s, where F is the force that we want to compute. We
only need a value for ∆t. A reasonable guess is ∆t = .2 s. Then F ∗.2 ≈ 143.75
or F ≈ 144/.2 = 720 lb.
Is 720 pounds more or less than you were expecting? On the one hand,
720 pounds worth of barbells would flatten most people in the gym. On the
other hand, you may have seen an episode of Sport Science in which sensors
measured a Ray Lewis hit on a crash test dummy at 2200 pounds. Factors
that can increase the calculated force are the size and speed of the player, the
impact time, and the velocity after the collision. In the Sport Science collision,
the dummy was driven backwards by Lewis. The next example picks up the
story at that point.
58 Sports Math

Example 4.2 A 160-pound crash dummy moving 14 mph is tackled by Ray

Lewis, who delivers a force of 2200 pounds for 0.1 s. At what speed is the
dummy moving right after the collision?
Solution. Convert the dummy’s speed to ft/s, so that v = 14(5280/3600) ≈
20.5 ft/s and m = 160/32 = 5 slugs. The dummy’s linear momentum before
the collision is p = 20.5(5) = 102.5 lb-s. The impulse for the hit is J = F ∆t =
2200(.1) = 220 lb-s. Since impulse equals the change in linear momentum, the
dummy’s linear momentum after the collision is 102.5 − 220 = −117.5 lb-s. If
the dummy’s velocity after the collision is denoted by w, then 5w = −117.5
or w = −23.5 ft/s, where the negative indicates that the dummy is moving in
the opposite direction. Lewis did not merely stop the dummy’s progress, he
knocked it in the opposite direction at a faster speed than it had been moving
forward! In this case, Lewis would need to wrap up the dummy with his arms
to complete the tackle, or Lewis and the dummy would fly apart.

In the first two examples, we assumed that a con-

stant force was applied for a period of time. In re-
ality, the force builds up from zero to a maximum
and then decreases back to zero. For a non-constant
force, the calculation of impulse requires some ge-
ometry and, in general, calculus. However, the un-
derlying concept is simple. Figure 4.1 shows a force FIGURE 4.1: Area
that lasts for 0.2 s, building linearly from 0 lb to a for Nonconstant Force
maximum of 1200 lb at time 0.1, then dropping lin-
early to 0. The impulse equals the area “under the
curve,” in this case the shaded triangle formed by
the graph and the horizontal x-axis. In Figure 4.2,
we have the same basic situation of a force building
up to a peak of 1200 pounds at time 0.1 s before
dropping back to 0 pounds. This time, the buildup FIGURE 4.2: Area
is non-linear (in fact, the graph is a parabola). We for Nonlinear Force
can argue in two different ways that the impulse for
this force is larger than that of the force in Figure 4.1. First, it should be clear
visually that the shaded area is larger in Figure 4.2 than in Figure 4.1. Second,
for any given time value the graph of the parabola is above the straight line
segments in Figure 4.1. This means that the force is larger, and therefore will
create a larger impulse.
Exact values can be computed in many cases.

Example 4.3 Compute the impulse for the forces in Figure 4.1 and Figure
4.2, and for a constant force of 1200 pounds for 0.2 s.
Solution. The impulses equal the areas under the curves, as shaded in the
figures. The area of a triangle is 1/2 base times height. In Figure 4.1, the height
is 1200 and the base is 0.2, so the area is .5(1200)(0.2) = 120 and the impulse
is 120 lb-s. In Figure 4.2, we have the area bounded by a parabola. This
 Collisions 59

calculation requires calculus (see the calculus box below) or Archimedes’

rule that the area is 2/3 times base times height; in this case, the area is
2
3 (1200)(0.2) = 160 and the impulse is 160 lb. For a constant function, the
graph forms a rectangle with area equal to base times height, or 1200(0.2) =
240 with an impulse of 240 lb-s. Note in this last case that we can use the
constant force formula F ∆t to get the impulse.

Before taking a quick look at the calculus of computing areas, we ask a

question. How should the forces in Example 4.3 be reported? They all reach
a peak of 1200 pounds, but their effects as measured by impulse are quite dif-
ferent. Therefore, saying that each is a 1200-pound force could be misleading.
One resolution is to compute the average force in each case. Since a constant
force of 600 lb applied for 0.2 s produces an impulse of 120 lb-s, we can say
that the average force in Figure 4.1 is 600 lb. Similarly, the average force in
Figure 4.2 is the impulse divided by the time span, or 160 lb-s / 0.2 s = 800
lb. This clearly indicates that this force has a bigger impact. However, if the
point is to impress the audience with the largest number possible, the peak
force of 1200 lb is more impressive than the average force of 800 lb.

Calculus Box: Integration

One of the most important techniques of one variable calculus is integra-
tion. In the case of a positive function f (x), the (definite) integral from x = a
to x = b gives the area under the graph of f , above the x-axis, and between
x = a and x = b. If we accept at face value that impulse is given by such
an area, then we can conclude that impulse is computed with an integral.
However, this avoids the obvious question of why impulse equals this area. It
is helpful to derive the formula for impulse from first principles.
For a constant force F over a length of time ∆t, impulse is given by F ∆t.
A force of 1200 lb for 2 s has an impulse of 1200(2)=2400 lb-s. Let’s make the
force non-constant by saying that for the first second the force is 800 lb and
for the second second the force is 1000 lb. The impulses are 800 lb-s and 1000
lb-s, which add to a total impulse of 1800 lb-s. Now, suppose that we have
different forces for each tenth of a second: F1 lb from time 0 s to time 0.1 s,
F2 lb from time 0.1 s to time 0.2 s, and so on through F20 lb from time 1.9 s
to time 2.0 s. Then the total impulse is
20
X
F1 (.1) + F2 (.1) + ... + F20 (.1) = Fi (.1)
i=1

We now generalize to a force F (t) on an interval a ≤ t ≤ b. Divide the interval

into n subintervals of equal length ∆t = b−a
n . If n is large, we expect that the
60 Sports Math

force F is approximately constant on each subinterval, so that F (ti )∆t is a

good approximation to the impulse on the i-th subinterval Si , with ti being a
point in Si . Then the impulse J is approximately
n
X
J≈ F (ti )∆t.
i=1

The larger n is, the better the approximation should be. In the limit,
n
X Z b
J = lim F (ti )∆t = F (t) dt
n→∞ a
i=1

so that the integral comes into play because we approximate the impulse as the
sum of impulses over small time intervals. In the limit, this sum approaches an
integral, which (because force is a non-negative quantity) equals area under
the curve.

Example 4.4 For the force F (t) = 120000t(.2 − t), 0 ≤ t ≤ 0.2, in Figure
4.2, compute the impulse.
Solution. By the above argument, the impulse is
Z 0.2
t=0.2
120000(.2t − t2 ) dt = 120000(.1t2 − t3 /3)t=0
0

= 120000 .1(.2)2 − (.2)3 /3 = 160 lb-s

 

as we previously obtained using Archimedes’ rule.

Giving to Receive
The impulse-momentum equation can be used to explain some basic tech-
niques in sports. If you have ever played catch, think for a moment about
how you catch a baseball. First of all, you want to use a glove. Further, you
probably have a nearly unconscious habit of pulling your hand back when the
ball arrives. Both are good ideas, as we see below.
To catch the ball, you must remove all of its linear momentum. This change
of linear momentum equals the impulse you apply to the ball (and, by New-
ton’s third law, the same impulse is applied back to you). Any combination of
force and time can accomplish this, as long as F ∆t equals the desired quantity.
The important point is this: if you can increase ∆t, the impact time, you can
decrease the force and achieve the same impulse. A padded glove slows the
ball down and lengthens the impact time. Pulling your hand back and giving
 Collisions 61

with the ball also increases the contact time. In both cases, the longer contact
time requires less force. Your hand appreciates the reduction in force.
A different aspect of the same principle occurs when a soccer player tries
to trap a long pass. If the player is able to “catch” the ball with a foot that
gives with the ball, the force applied to the ball is reduced. Then the change
in linear momentum is reduced from a large change that sends the ball in the
opposite direction to a smaller change that brings the ball down with zero
velocity.

Tendons and Tennis

Impulse can be computed from the graph of
force as a function of time. A different graph,
also of interest in sports, shows force as a func-
tion of distance. Our feet endure constant com-
pression and stretching as we move around. The
ground pushes the balls of our feet, while our
Achilles tendons pull on the backs of our feet. A
specific force, shown on the vertical axis, causes FIGURE 4.3: Force versus
a specific compression of the foot, shown on the Displacement
horizontal axis. Figure 4.3 is based on data from
Alexander’s Exploring Biomechanics, where the
foot compression is measured in millimeters and
the force is measured in kilo-Newtons.
After the foot is compressed by these forces,
it springs back to normal length. In doing so,
it pushes back against the ground and Achilles
tendon. Alexander also measured these forces,
which form the lower of the two curves in Figure
4.4. Notice that the forces during relaxation of
the foot are smaller than the forces during com-
pression. The reduction in force is a product of FIGURE 4.4: Force versus
the loss of energy to heat in the foot. The area Displacement
under each curve is proportional to the energy
change. If Ac is the area under the compression
curve and Ar is the area under the relaxation curve, then the ratio Ar /Ac
is the proportion of energy retained in the compression/relaxation cycle. In
other words, this is the efficiency of the foot at retaining energy. Alexander
measured this value at about 78%. By comparison, tendons from the hind
legs of wallabies retain about 93% of their energy, making wallabies better
jumpers than humans.
This analysis also applies to collisions in which balls compress and relax.
62 Sports Math

When a tennis ball hits the strings of the tennis racket, the ball flattens al-
most completely before popping back out to its spherical shape. The same
phenomenon happens, to a lesser extent, to golf balls and baseballs in their
collisions with clubs and bats. The proportion of energy lost varies from sit-
uation to situation. For collisions of balls with other objects, we redefine the
energy loss in a convenient way.

Coefficient of Restitution
Think of dropping a ball from some height. It falls to the floor and then
bounces back up, but does not make it back to its original height. The next
bounce is lower still. It turns out that there is a regularity to the bounces.
If the ball reaches 70% of its original height on the first bounce, the second
bounce will be about 70% of the first bounce (now 49% of the original height).
On each bounce, the ball retains about 70% of its energy.
Recall that the height that a ball is dropped from determines its speed
when it hits the ground, and the speed at which a ball is launched from the
ground determines its height. Since the heights are different, it must be that
the speed vb at which the ball hits the ground is greater than the speed va at
which it is launched back into the air. The ratio of the speeds is named the
coefficient of restitution (COR). That is,
va
COR = .
vb
The definition can apply to any collision of two objects, with vb equalling
the relative velocity before the collision and va the relative velocity after the
collision. For example, if right before the collision a bat is moving 80 mph
in one direction (the negative direction) and a ball is moving 90 mph in the
opposite (positive) direction, then vb = 90 − (−80) = 170 mph. If the bat and
ball exit the collision moving in the same direction with ball speed 120 mph
and bat speed 35 mph, then va = 120 − 35 = 85 mph, and the COR for this
85
collision is 170 = 0.5.
Notice the wording of the last statement. It is important to realize that
COR is a property of the collision of two objects and not just one object.
Therefore, it does not make sense to ask for the COR of a baseball. You will
get a higher bounce (higher COR) bouncing a baseball off wood than off a
pillow. For brevity, we might say that the COR of a baseball is 0.546, but
this is only valid for certain types of collisions. (In the case of major league
baseballs, COR is measured for balls bouncing off of wood at 85 ft/s.) We
sometimes assume, given no other data, that the COR for other types of
collisions is approximately the same. However, the COR depends on many
factors, including the speeds of the objects involved and their composition.
 Collisions 63

From the definition, it should be clear that COR is a number between

0 and 1. If COR = 0, then the collision is called perfectly inelastic and the
objects stick together; an example is a football tackle. If COR = 1, then the
collision is called elastic and no energy is lost. Our interest will be in inelastic
collisions with 0 < COR < 1. In general, the larger the COR is the bouncier
or livelier the collision is. This has important ramifications in almost all ball
sports.

Incoming and Outgoing

The situation to be analyzed in this section applies to many sports. To
make the discussion concrete, imagine a baseball being hit by a bat. Before
the collision, the ball and bat move in opposite directions with speeds vball
and vbat , respectively. After the collision, the ball and bat move in the same
direction with speeds wball and wbat , respectively. The situation is depicted in
Figure 4.5.

FIGURE 4.5: Speeds Before and After

Two principles and some algebra give us some insight into the interactions
of ball and bat. The first is COR, which in this context is given by
wball − wbat
COR = (4.1)
vbat − −vball
where the double negative in the denominator is due to the ball moving in the
negative direction. The second principle is conservation of linear momentum,
which in this case means

mbat vbat − mball vball = mbat wbat + mball wball (4.2)

where we again account for the direction of the moving ball.

Typically, we want to know how the ball speed after the collision depends
on the other factors. For example, how can we give the ball enough speed
to clear the fence for a home run? We can solve equation (4.1) for wbat and
64 Sports Math

substitute into equation (4.2). After some rearrangement, we get

mbat (COR vball + (1 + COR) vbat ) − mball vball

wball = . (4.3)
mbat + mball
This looks complicated, but it is actually easy to use and interpret.

Example 4.5 For a baseball of weight 5.25 oz and speed 90 mph, a baseball
bat of weight 32 oz and speed 80 mph, and a COR of 0.4, find the speed of
the ball off the bat.
Solution. From equation (4.3), the speed is

32(0.4 · 90 + 1.4 · 80) − 5.25 · 90

wball = = 114.5 mph.
32 + 5.25
Explanations are in order. We did not convert weight to mass, or speeds to
ft/s. An examination of the units shows why. COR is unitless, so wball =
32 oz (0.4·90 mph+1.4·80 mph)−5.25 oz·90 mph
(32+5.25) oz . The “oz” units cancel, as any unit of
mass would have. That leaves units of “mph.” This is nice! As long as we’re
consistent, the units do not matter.

Easier than you expected, right? Unfortunately, we still have some issues
to resolve. The batted ball speed of 114 mph is on the high side for the major
leagues, but the pitch speed of 90 mph and bat speed of 80 mph are on the
low side. An assumption underlying equation (4.3) is that the ball and bat
collide in a line through their centers of mass. A fly ball is produced by the bat
hitting slightly below the ball’s center of mass, so our calculation only holds
for a line drive with no spin. More commonly, some of the energy transfer
from the bat to the ball is diverted from ball speed to ball spin. Further, the
major league specification for a baseball is that the COR must be between
0.514 and 0.578. The value of 0.4 used in Example 4.5 is too low. I chose it, to
be honest, so that the calculated ball speed was low enough to be plausible.
To justify this change, note that the major league testing procedure is to fire
a ball at 85 ft/s at a wooden wall. The relative speed of ball and bat before
the collision in Example 4.5 is 170 mph, about 250 ft/s. The large difference
in laboratory and playing field velocities makes it likely that the official COR
of about 0.55 is too large.
In Example 4.5, we used a bat weight of 32 oz. This is close to the average
weight in the major leagues, but think about the physical situation. The bat
doesn’t swing itself at 80 mph. A person is attached to the bat. Shouldn’t we
include the mass of the batter as well as the bat? The answer is that the mbat
term in equation (4.3) is actually a new quantity called effective mass. The
effective mass of the bat depends on where on the bat the ball is hit. If it is
hit in line with the center of mass of the bat, then effective mass and mass of
the bat are the same. Generally, on off-center hits, the effective mass of the
bat is less than the mass of the bat.
 Collisions 65

Brody, Cross, and Lindsey introduce detailed information about tennis

rackets in The Physics and Technology of Tennis. In the racket-ball collision,
the effective mass issue is especially confusing, since the strings are the only
part of the racket that touch the ball. The center of mass of the racket is
near the neck, barely on the strings. The racket’s maximum COR occurs at
this spot, and effective mass equals racket mass here. Moving up the racket
reduces effective mass. By the time you reach the center of the racket, the
effective mass is reduced to about half the mass of the racket.
An interesting experiment illustrates the role of grip on effective mass.
Stand a tennis racket on its end and throw a ball at the center of the strings.
The racket will be knocked over, but mark where the ball first bounces. Then
repeat the process with someone holding the racket tightly by its handle. The
ball will bounce to the same spot! The ball speed is not affected at all by the
complete absence of a grip! The effective mass does not depend on human
interaction: the strings stretch and then send the ball on its way before the
vibrations of the racket can reach and return from the hand. The ball has left
before the strings know whether or not your hand is there.
It has been shown that the effective mass for a football field goal kicker is
larger if the kicker kicks from the side, soccer-style, rather than head on. This
is one reason that all kickers, even those who never played soccer, kick from
the side.

Derivative Works
What effect does a decrease in effective mass have? These and other basic
questions can be answered with some algebra and/or calculus. For example,
take equation (4.3) and think about the effect of increasing mbat . Since mbat
multiplies a positive constant in the numerator, the numerator will increase.
But so will the denominator. We need calculus here, in particular the deriva-
tive. As is shown in the calculus-based exercises, an increase in the effective
mass of the bat or racket always results in an increase in ball speed. This
should coincide with your experience. Swing a big stick to hit it hard!
It is easier, mathematically, to see what happens if vbat is increased. Since
vbat is found only in the numerator multiplied by a positive constant, an
increase in vbat results in an increased ball speed. This should make sense:
a faster swing has the potential for the hitting the ball harder (if you can
make contact). An interesting analysis involves changes to vball . Since vball is
multiplied by mbat COR − mball , an increase in vball increases the ball speed if
mbat COR − mball > 0, but decreases the ball speed if mbat COR − mball < 0.
In baseball and most sports, mbat COR − mball > 0, so the faster the ball
comes in, the faster it goes back out.
66 Sports Math

The Way the Ball Bounces

One of the most common and important collisions in sports is one between
the ball and ground. The mechanics here are more complicated than in equa-
tion (4.3), because the ball typically moves at an angle to the center of mass of
the striking object (Earth). However, we can determine some basic principles
from Newton’s laws.
In the figure, a ball hits the ground while moving
to the right and downward with no spin. When it hits
the ground, the ball pushes to the right and down-
ward on the ground, which (by Newton’s Third Law)
pushes the ball upward and to the left. The leftward
push causes the ball to spin in a direction that we recognize as topspin.
The ball leaves the ground with topspin, at an
angle that is different from the incoming angle. The
overall speed of the ball has been reduced, but the
upward and leftward push of the ground gives the
ball relatively more vertical motion and less horizon-
tal motion; the angle to the ground will be greater outgoing than incoming.
The amounts that speed and angle change are determined by the surface prop-
erties of ball and ground. The greater the coefficient of friction, the longer the
ball and ground stay in contact and the more effect the ground has.
If the incoming ball has backspin, the bottom of the ball in the above figure
is moving left-to-right. The backspin ball then pushes the ground harder to
the right than the no-spin ball, and receives a harder push to the left in return.
If the pushback to the left is hard enough, the ball can bounce backwards.
You have probably seen this effect with a tennis ball, ping pong ball, baseball,
or golf ball. Until 2010, you rarely saw this in football, but punters have now
started putting backspin on the ball (by kicking it with the nose down) to
keep punts from bouncing into the endzone.
Backspin can explain the phenomenal success of Rick Barry shooting free
throws. One of the all-time greats in the NBA and ABA, in his last eight
seasons Barry made 2496 out of 2731 free throws (91.4%!) while leading the
league in percentage made in six of the eight years. Find a video of Barry
shooting free throws: he used an underhand motion! One of the advantages of
this technique is the ability to put extra backspin on the ball. If the ball hits
the rim, the backspin will result in a softer bounce that increases the chances
that the ball goes in.
One aspect of the backspin bounce may seem counterintuitive to tennis
players. The stronger leftward push of the ground takes more horizontal ve-
locity away from a backspin ball than a no-spin ball, so that the angle of the
bounce is greater than that of the no-spin ball. Tennis players hit slices with
lots of backspin to keep the ball low. How does this work if backspin makes
 Collisions 67

the ball bounce higher? The above graphics have confused the issue. Because
backspin creates an upward Magnus force, the trajectory of the ball can be
much flatter than a no-spin shot, so that its incoming angle can be much
smaller than that of a typical shot. The small incoming angle is what keeps
a backspin slice low, even though it bounces at a higher angle than a no-spin
shot with the same incoming angle.
Tennis is played on many different types of surfaces: grass and clay are the
old traditionals, but a variety of synthetic courts allow tournaments to control
the pace of play. How can we measure the pace of a tennis court? You might
start with COR; if the ball loses most of its speed as the result of a small
COR, then the court must be slow, right? Unfortunately, grass courts have
much smaller COR’s than do clay courts. (Think about it: would you expect
a ball to bounce higher off of grass or clay?) This contradicts our knowledge
that grass is fast and clay is slow. Another factor that influences court speed
is the coefficient of friction. This measures the resistance to motion while
the ball is in contact with the ground. Grass is slick with a low coefficient
of friction, while clay is rough with a high coefficient of friction. A quantity
named Court Pace Rating combines the two coefficients (COR and friction)
into a meaningful rating.

Freeze Frame
One of the most interesting revelations of
sports high-speed photography is the chore-
ography of a hockey slap shot. The player
takes a big windup and slaps the stick into
the ice right behind the puck, with one hand
positioned near the bottom of the stick. The
low positioning of the bottom hand reduces
the velocity of the stick at impact, but the
power of the slap shot comes more from po-
tential energy than kinetic energy. Notice the
extreme curve of the stick from top hand to
bottom hand in the picture.
The bottom hand pressing down on the
stick causes it to bend substantially, storing
potential energy. As the stick flexes back into
shape, the puck is swept along. In the pho-
tograph to the right, notice how far along
the stick and puck are without separating.
The stick’s potential energy is transferred to
the puck as kinetic energy. A Sport Science
68 Sports Math

episode measured a shooter’s arm speed at 75 mph, stick speed at 80 mph, and
puck speed at 100 mph. Photographs have shown the stick and puck making
contact multiple times. One problem with an extreme bending of the stick is
that too much bending will cause the sticks to break. Phil Kessel of the Maple
Leafs and other players required specially made sticks with extra strength.
Modern sticks are made with flexible lightweight carbon materials and are
designed for optimal flex. Manufacturers proclaim that their sticks will “load
up” on slap shots, consistent with our analysis of the stick-puck collision.

Exercises
In these exercises, T refers to thinking problems, conceptual problems re-
quiring no calculations. C refers to problems requiring calculus or significant
computer calculations. P refers to projects; these are ideas for further inves-
tigation (hints and resources are at the book’s web site).

4.1 A 220-pound running back moving 18 mph runs into a 320-pound defensive
lineman. (a) What is the speed needed for the lineman to stop the running back?
(b) Assuming that the collision lasts 0.2 second, what is the average force?
(c) Find the peak force if the force is piecewise linear as in Figure 4.1.
(d) Find the peak force if the force is parabolic as in Figure 4.2.
4.2 A 180-pound defensive back moving 18 mph runs into a 240-pound running
back moving at 12 mph. (a) What is the outcome of the collision?
(b) Assuming no other forces, what is the combined velocity of the pair after the
collision?
(c) Assuming that the collision lasts 0.2 second, what is the average force?
(d) Find the peak force if the force is piecewise linear as in Figure 4.1.
(e) Find the peak force if the force is parabolic as in Figure 4.2.
4.3 Find the impulse for each force. (F is in pounds, and t is in seconds.)
(a) F (t) = 100 for 0 < t < .1; F (t) = 300 for .1 ≤ t ≤ .2; F (t) = 600 for
.2 ≤ t ≤ .4; F (t) = 300 for .4 ≤ t ≤ .5; F (t) = 100 for .5 ≤ t ≤ .6;
(b) F (t) = 200t for 0 < t < 2; F (t) = 1200 − 400t for 2 < t < 3
(c) F (t) = 400t(3 − t) for 0 < t < 3
(d) F (t) = 600t(2 − t) for 0 < t < 2
4.4 Given the following compression/relaxation profiles for tendons (see Figure
4.4), which one represents the more efficient tendon? If the curves instead are
for two tennis balls, which ball is livelier?
(a) 21 (x2 + x) and x2 for 0 < x < 1 (b) 21 (x2 + x) and x4 for 0 < x < 1
4.5 The following statements give official regulations for sports equipment. For
each, give the speed of impact, how the speed at impact compares to typical
collision speeds in that sport, and the range of CORs.
(a) A basketball dropped from a height of 1.8 m bounces to a height between
1.2 m and 1.4 m.
 Collisions 69

(b) A lacrosse ball dropped from a height of 72 in bounces to a height between

43 in and 51 in.
(c) A tennis ball dropped from a height of 100 in bounces to a height between
53 in and 58 in.
(d) A high-altitude tennis ball dropped from a height of 100 in bounces to a
height between 48 in and 53 in.
4.6 To control the “trampoline effect” of hollow metal drivers, golf now legislates
the COR of a driver-ball collision to be no more that 0.83. If a driver is moving
at 140 mph, find the maximum relative speed of the ball and driver after the
collision. Is this the same as the maximum ball speed?
4.7 Show that COR = ab where a is the height of the bounce of a ball dropped
p

from height b.
4.8 (a) For the ball and bat of Example 4.5, find the ball speed wball if the bat
speed is increased to 81 mph.
(b) For the ball and bat of Example 4.5, find the ball speed wball if the incoming
ball speed is increased to 91 mph.
(c) Which has the larger impact on ball speed, bat speed or ball speed?
4.9 (a) A 57 gm tennis ball moving 60 mph is hit by a 300 gm racket moving 80
mph in the opposite direction. If COR = 0.74, find the ball speed after impact.
(b) A golf ball of mass 0.05 kg at rest is hit by a golf club of mass 0.17 kg moving
120 mph. If COR = 0.8, find the ball speed after impact.
4.10 Solve equations (4.1) and (4.2) for wbat .
4.11 Use the solution of exercise 4.10 to determine whether wbat increases or
decreases when (a) vball increases; (b) vbat increases.
4.12 Find wbat in exercise 4.9, parts a and b.
4.13 T Use the concept of impulse to explain good techniques for each:
(a) catching a football; (b) landing from a large height;
(c) dribbling a basketball; (d) hitting a drop volley in tennis.
4.14 T Use the concept of impulse to explain why running in sand is more
difficult than running on concrete.
4.15 T Tennis rackets can be strung at different tensions. Given that looser
strings create longer impact times, discuss whether looser or tighter strings would
produce (a) more ball speed; (b) more control.
4.16 T Discuss what would happen to the effective mass of a field goal kicker
if his planting foot slipped.
4.17 T The effective mass of a tennis racket does depend on the tightness of
the player’s grip if the ball is hit off-center (not on the line running through the
handle and out the top of the racket). Explain.
4.18 T It was noted in the text that wball decreases with an increase in vball if
mball > mbat ·COR. Explain in physical terms why this makes sense.
4.19 T Two balls with positive horizontal (and negative vertical) velocity hit
the ground at the same angle. Ball A has no spin, and ball B has topspin. Which
ball bounces at a larger angle to the ground? Explain.
4.20 T Is it possible for a ball with topspin to have a larger horizontal velocity
after hitting the ground than before? Explain. Discuss how this result applies to
ground balls in baseball and ground strokes in tennis.
70 Sports Math

4.21 T A ball moves horizontally with backspin before hitting a wall. Does the
spin cause the ball to rebound off the wall higher or lower than a ball with no
spin? What type of spin does the ball have after hitting the wall? Discuss how
this result applies to shooting a basketball off the backboard.
4.22 T Repeat exercise 4.21 with topspin. Discuss how the result applies to a
tennis volley of a topspin shot.
4.23 T A baseball player hits a fly ball by making contact with the top portion
of the bat barrel hitting the bottom part of the ball. What type of spin is
produced? Does this spin increase or decrease the distance of the fly ball?
4.24 T Typically, a fastball has more speed than a curveball. Only taking this
into account, explain why it should be easier to hit a home run off of a fastball.
For which pitch would a fly ball have more backspin? Simulations have shown
that a well-struck hit off of a curveball will travel farther than off of a fastball.
Briefly explain how this could be true.
4.25 T A hockey slap shot starts with the player taking a big swing and hitting
the ice right behind the puck. Explain why a similar technique on the golf course
would not be effective. Use spin to explain why golfers get better results hitting
the ball directly.
4.26 T Compared to grass and clay, describe how large the COR of a basketball
court would be for a tennis ball. How large would its coefficient of friction be?
How large would its Court Pace Rating be?
4.27 T In the Sport Science episode “Human Flight” Jerry Rice talks about
catching footballs with the fingertips: “If it hits you in the palms of your hands
you’re going to have that ricochet.” Discuss the advantage of fingertips over
palms in terms of impulse.
4.28 C Compute the impulse and average force of each force. (a)
1, 000, 000t2 (0.2 − t) for 0 < t < 0.2. (b) 1000 sin(πt/0.2) for 0 < t < 0.2.
4.29 C Prove Archimedes’ Rule for F (t) = at(b − t) for 0 < t < b with a > 0
and b > 0.
4.30 C For a constant force F applied over a distance d, work is defined by
W = F d. A force F (x) is applied at location x for a ≤ x ≤ b. Derive the more
Rb
general formula W = a F (x)dx.
Rb
fc (x) − fe (x)dx
4.31 C Use the data to estimate the efficiency 1 − a R b .
f (x)dx
a c
x in 0.0 0.1 0.2 0.3 0.4
(a) tennis ball data fc lb 0 25 50 90 160
fe lb 0 23 46 78 160
x in 0.0 0.1 0.2 0.3 0.4
(b) baseball data fc lb 0 250 600 1200 1750
fe lb 0 230 450 700 1750
x mm 0.0 0.75 1.5 2.25 3.0
(c) Wallaby tendon data fc N 0 110 250 450 700
fe N 0 100 230 410 700
x mm 0.0 2.0 4.0 6.0 8.0
(d) human foot data fc N 0 300 1000 1800 3500
fe N 0 150 700 1500 3500
 Collisions 71

4.32 C Use equation (4.3) to compute the derivative of wball with respect to
mbat . That is, treat mbat as the variable and all other parameters as constants.
Show that the derivative is positive, and interpret this to mean that a bigger bat
will hit a ball harder.
4.33 C Compute the derivative of wball with respect to mball , determine if it is
positive or negative, and interpret the result.
4.34 C Use the solution to exercise 4.10 to compute and interpret the derivative
of wbat with respect to (a) mbat ; (b) mball
4.35 C In Example 4.5, wball ≈ 0.2vball + 1.2vbat . Explain why the increase in
wball in exercise 8(a) is 1.2 mph and the increase in exercise 8(b) is 0.2 mph.
148mbat − 90mball
4.36 C In Example 4.5, wball = .
mbat + mball
(a) Use this formula to compute the derivative of wball with respect to mbat and
interpret the result in terms of the change in wball for a change in mbat .
(b) Repeat part (a) with respect to mball .
4.37 P Drop a ball from different heights and construct a graph of COR as a
function of impact velocity. Does the graph appear to be linear?
4.38 P Bounce balls off of different places on a tennis racket to determine COR
as a function of location on the racket.

Further Reading
Peter Brancazio’s Sport Science is an excellent introduction to the physics
of sports. The bits on catching a ball, throwing a ball at a tennis racket, and
the effective mass of a field goal kicker are from Brancazio.
Robert Watts and Terry Bahill’s Keep Your Eye on the Ball is the best of
several physics of baseball books.
A search for “Physics of (favorite sport)” will bring up many results.
I can recommend The Physics of Basketball (Fontanella), The Physics of
Hockey (Hache), Football Physics (Gay), Newton’s Football (St. John and
Ramirez), The Science of Soccer (Wesson), The Physics and Technology of
Tennis (Brody, Cross, and Lindsey), Golf Science (Smith), The Science of
Golf (Wesson), Golf By the Numbers (Minton), and The Physics of NASCAR
(Leslie-Pelecky).
Football Physics is the source of impact time for a football collision.
Several episodes of John Brenkus’s Sport Science are available at
http://espn.go.com/espn/sportscience/. Accessed 8-11-2015. The Ray Lewis
“Block and Tackle” segment was referenced in the text. All segments are in-
formative and entertaining.
I am, of course, partial to Smith and Minton’s Calculus as a calculus
reference. It has numerous sports problems.
R. McNeill Alexander’s Exploring Biomechanics is the source for the in-
formation about wallaby tendons and human feet.
72 Sports Math

Before co-authoring The Physics and Technology of Tennis, Howard Brody

published Tennis Science for Tennis Players.
Austin Murphy’s article “The Nail in the Coffin” about Aussie-style punt-
ing with backspin appeared in the October 27, 2014 issue of Sports Illustrated.
Lee Torrey’s Stretching the Limits has a nice description of the
hockey slap shot. Several excellent bent-stick pictures are online. A
search for “physics of hockey slap shot” will turn up good sites such
as http://physicsofhockeyproject.weebly.com/shooting.html. Accessed 8-11-
2015.
A good article from the Wall Street Journal about hitting curveballs far-
ther than fastballs is at http://www.wsj.com/articles/SB108568923118723199.
Accessed 8-11-2015.
Chapter 5
Ratings Systems

Introduction

From 1998 to 2013, the participants in college football’s national cham-

pionship game were partially determined by computer ranking systems. The
use of computers by the Bowl Championship Series (BCS) system was often
ridiculed by football reporters. However, there was an explosion of ranking
systems during the BCS era. The above graphic is a screen capture from the
excellent Massey Ratings web site, showing some of the 124 ranking systems
submitted during the week of November 9, 2013. Presumably, this is just the
tip of the rankings iceberg.
In this chapter, we develop several matrix-based systems for rating sports
teams or individuals. We will address a number of questions. What is the goal
of a ranking system? What is the difference between a rating and a ranking?
Are computer systems complicated? How is strength of schedule incorporated
into the systems? How can ranking systems be evaluated?

73
74 Sports Math

Right versus Best

Imagine that you are in charge of ranking teams in your favorite sport.
Are you more interested in picking the “right” teams or the “best” teams?
The distinction may be unclear at first, but consider the following two teams.
Team A has the best record in the league. However, its schedule was easy and
it often came from behind to post narrow victories. Team B has the fifth best
record in the league, against the toughest schedule in the league. Its losses were
all early in the season with several players injured. With everybody healthy,
team B finished the season winning its last several games by wide margins.
The records make A the “right” number one, but the hot streak at the end
makes B the “best” team.
Most leagues design ranking systems that emphasize the choice of the right
team. Won-lost records dominate, and tie-breakers emphasize head-to-head
matchups. This makes sense when the teams’ schedules are equal or nearly
equal. College sports teams can have wildly different strengths of schedule.
Few people would choose an undefeated team from a minor league over a
once-defeated team from a major league. Yet college football rankings often
follow a “check-out line” mentality. If you are first in line, nobody is allowed
to move ahead of you unless you lose. If you lose, the teams behind you all
move up one slot in an orderly fashion. This type of ranking choreography
largely ignores the issue of which team is actually the best.
A slightly different way of phrasing the question is to ask whether you are
trying to evaluate the past or predict the future. Take teams A and B from
above. If I seed team A #1 and team B #2 and then predict team B to win
the game, it is clear that my ranking system is more about picking the right
team based on past records than picking the best team based on who will win.
Here is the point: a mathematically sound analysis starts with clear objec-
tives and finishes with a precise evaluation of how well the objectives are met.
A major distinction between computer ranking systems and human ranking
systems is that we humans rarely have a single set of objectives. We some-
what randomly balance the right choices with the best choices, sometimes
using our predictions of the future but often not. An understanding of our
basic irrationality is important when evaluating computer systems.

Ratings versus Rankings

The terms “rating” and “ranking” have thus far appeared in what may
seem to be an interchangeable fashion. However, they are different. A rating
system assigns a number to each team that corresponds to its strength. A
 Ratings Systems 75

ranking system lists all of the teams in an order that corresponds to their
strengths. A rating system can easily be turned into a ranking system by
listing the teams in numerical order, but a rating system may provide extra
information, such as how closely matched the teams are.
We assume that the ratings and rankings are transitive. That is, if we
claim that A is better than B and B is better than C, then we implicitly
acknowledge that A is better than C. This may seem obvious, but we will
see in Chapter 6 that the issue is complicated. For now, note that our rating
systems are one-dimensional in their output, producing a single number that
characterizes each team’s strength. We will not, for example, incorporate styles
of play that could lead to a conclusion that while A will beat B and B will
beat C, C matches up well with A and would beat A.
We will introduce three rating systems, two of which were part of the BCS
computer ranking system. To get an idea of what a “computer system” is like,
here are the main assumptions for each of the systems.
Massey System: a team’s rating is the average of its opponents’ ratings
plus its net points per game.
1+w
Colley System: a team with w wins in g games has a rating of ,
2+g
with w modified for strength of schedule.
Elo System: a team’s rating is updated after each game, increasing after
a win and decreasing after a loss, with the size of the change depending on
the quality of the opponent.
The only intimidating aspect of any of these systems is the amount of time
it would take to calculate the ratings by hand. That is why they are computer
systems: computers are needed to do a sizable amount of number crunching.

The Massey System

The Massey System is named for Kenneth Massey, who developed it as
an undergraduate student. Massey followed in the footsteps of others. The
version presented here was published as a UMAP Module, and is a slight
modification of work by Jech and others. It is interesting that this is the
system that bears Massey’s name, since in graduate school Massey developed
a more sophisticated rating system that was part of the BCS system, and
which he now applies to a variety of sports at the Massey Ratings web site.
Objective: The difference in ratings of two teams represents the
point differential in a contest between the two teams.
This is a rating system, so each team is assigned a number representing its
strength. The unit of measurement is points, and the model is based on the
idea that if Dallas is rated 100 and Washington 88, then Dallas is 12 points
76 Sports Math

better. The rating could be used to predict the outcome of a game between
Dallas and Washington, or to rank Dallas above Washington.
Model: A team’s rating is the average of its opponents’ ratings
plus its net points per game.
Thus, if Dallas has outscored its opponents by a total of 80 points in
16 games (an average of 5 points per game), its rating is the average of its
opponents’ ratings plus 5. To see how this plays out, we work this simple
example.
Early in the season, a four-team Team W L PF PA
league has the results shown to the side. A 5 0 72 50
“W” is wins, “L” is losses, “PF” is points B 2 2 44 44
for, and “PA” is points against. At first C 2 2 40 40
glance, you would think that A is the D 0 5 40 62
best team, D is the worst, and teams B
and C are equal. This is not true, as we TABLE 5.1: Four-Team League
will see after getting some valuable in-
formation. What is missing?
We need to know which games have A B C D
been played, so that we can take into A - 2 0 3
account strength of schedule. Take the B 2 - 2 0
schedule shown to the side. This tells us C 0 2 - 2
that A and B have played twice, A and D 3 0 2 -
C have not played, A and D have played
3 times, and so on. Now, why should we TABLE 5.2: Games Played
rank B above C?
Team B played the best team in the league, while C did not. B had a
tougher schedule, so its 2-2 record is more impressive than C’s 2-2 record
against inferior competition (two games against D, which B did not play). If
you’re not convinced, think through the results and note that since B lost
twice to A, it must be that B won both games played against C!
The model for the Massey system translates to one equation for each team.
Name the team ratings a, b, c, and d. Team A scored 22 points more than its
opponents in 5 games, so team A’s net points per game is 22/5 = 4.4. Then
a equals the average of A’s opponents’ ratings plus 4.4: a = b+b+d+d+d
5 + 4.4.
Multiply by 5 to get 5a = 2b + 3d + 22 or 5a − 2b − 3d = 22. Team B’s net
points per game is 0/4 = 0 so its equation is b = a+a+c+c
4 or 4b − 2a − 2c = 0.
In alphabetical order, the complete set of equations is

5a − 2b − 0c − 3d = 22

−2a + 4b − 2c − 0d = 0
−0a − 2b + 4c − 2d = 0
−3a − 0b − 2c + 5d = −22
Notice how the schedule and results are encoded in the equations. For a given
 Ratings Systems 77

team’s equation, its (unknown) rating is multiplied by its number of games

played. Subtracted from that are the other teams’ ratings, multiplied by the
number of games played against that team. On the right-hand side of the
equation is the net points scored (total, not per game). Easy!
All that’s left is to solve the equations.  
We do so by row reducing the corresponding 5 −2 0 −3 22
matrix. This is covered in a linear algebra  −2 4 −2 0 0 
course and (often) in precalculus. The ma-  0 −2 4 −2 0 
trix corresponding to the above equations is −3 0 −2 5 −22
shown to the right. Each row of the matrix corresponds to one team’s equa-
tion, and the entries in the rows are the coefficients of a, b, c, and d (in order!),
followed by the number on the right-hand side. Notice again that the number
of games against each team and the net points are readily apparent. Also,
notice that each column sums to 0. This will cause us some grief shortly.
The reduced matrix is shown to the right. Each 1 0 0 −1 6
row can be translated back into equation form. The 0 1 0 −1 4
top row translates to 1a + 0b + 0c − 1d = 6 or simply  0 0 1 −1 2

a − d = 6. Similarly, we have b − d = 4, c − d = 2 and 0 0 0 0 0
0 = 0. At one level, that is hugely disappointing:
we do not have a unique solution for the ratings. (This is a consequence of
the original columns summing to 0, which means that the rows are linearly
dependent.) Pick a value for d (any value) and you can find corresponding
values for a, b, and c.
One way to obtain a unique solution is to add an additional requirement,
such as the ratings summing to zero. However, our original objective was to
find the difference between any two teams’ ratings. The first equation tells us
that A is 6 points better than D. Subtracting the first two equations (a − d −
(b − d) = 6 − 4) tells us that A is 2 points per better than B. Similarly, A is
4 points better than C. This is true for any and all choices we make for d. So
all of the infinity of ratings give the same ranking and point differentials. In
particular, choosing d = 0 gives us c = 2, b = 4, and a = 6. Thus, the numbers
in the far right column of the reduced matrix can be used as ratings!

Connected Schedules
The results from this example are typical. The reduced matrix has a row
of 0’s and the next-to-the-last column of the matrix is otherwise all −1’s.
The numbers in the last column can be used as Massey point ratings. The
exception to this pattern is when the schedule is not connected.
In the above example, we have a basis for comparing teams A and C even
though they have not played: A played B, which played C. Teams A and C are
connected, even though they did not play. A schedule is connected if you can
put together a chain of games connecting any two teams: A played B, which
played C, which played .... It does not matter who won these games.
As long as the schedule is connected, the matrix will reduce as above and
78 Sports Math

we get our ratings. Most leagues’ schedules (even college football or basketball)
become connected after a small number of games.
Applying the Massey rating system
to 250 Division 1 college football teams
in the 2014 regular season, we get the 1 Alabama 69.2
top five shown to the right. This list, 2 TCU 68.4
unfortunately, fails the eye test for rea- 3 Oregon 65.2
sonableness. Undefeated Florida State is 4 Georgia 65.1
nowhere to be found (they were #19). 5 Ole Miss 64.7
How could the only undefeated team in
TABLE 5.3: Top 5 Points
the country not make the top ten? The
answer is to think about how the system works. It uses points and schedules,
period. If we want the ratings to take into account wins and losses, we need
to incorporate wins and losses into the system. The value of a rating system
depends on its underlying objectives!

Massey Win Ratings

To have the Massey system use wins and  5 −2 0 −3 5 
losses, we can simply replace net points with −2 4 −2 0 0
net wins (wins minus losses). In effect, we   0 −2 4 −2 0 

declare each game to have a 1-0 outcome. −3 0 −2 5 −5
Then proceed as before, reducing the matrix
and reading off the win ratings from the far-right column. The matrix for our
four-team league is shown to the right.
The reduced matrix to the right gives us the win 1 0 0 −1 15/11
ratings, in this case a = 15/11, b = 10/11, c = 5/11, 0 1 0 −1 10/11
and d = 0. The ranking of A,B,C,D is the same as  0 0 1 −1 5/11 

from the points ratings. However, it is not obvious 0 0 0 0 0
what the values of the ratings represent. Properly
scaled, the win ratings can be interpreted as points ratings. Multiplying by
22/5, the ratings become a = 6, b = 4, c = 2, and d = 0. More typically, the
win ratings are distinct from the points ratings.
The top five in win ratings from the
2014 college football regular season is 1 Florida St. 11.8
quite different from the points ratings 2 Alabama 11.8
seen earlier. Notice, in particular, that 3 Oregon 11.7
unbeaten Florida State is now number 4 Ohio State 11.7
one. While the points ratings penalized 5 TCU 11.6
Florida State for not winning its games
by large margins, the win ratings reward TABLE 5.4: Top 5 Wins
its undefeated season.
 Ratings Systems 79

Offense and Defense

A nice feature of Massey’s rating system is that the points ratings can be
split into separate ratings for offense and defense.
Model: A team’s points rating equals the sum of its offensive and
defensive ratings.
In a game between teams A and B, team A’s offensive rating minus team
B’s defensive rating gives the expected number of points for team A. Suppose
that team A has a points rating of 64 and team B is rated at 54. On a neutral
field, then, team A would be predicted to win by 10 points. Now, suppose that
team A’s rating breaks down as 40 on offense and 24 on defense (adding to
64), and team B’s ratings are 38 on offense and 16 on defense. Team A should
score 40 − 16 = 24 points to team B’s 38 − 24 = 14 points.
Recall that the equations for the points
(or wins) ratings do not have a unique 1 TCU 56.1
solution; we have arbitrarily set the final 2 Baylor 55.7
team’s rating to 0 in our examples. This 3 Oregon 55.0
lack of a unique solution carries over to 4 Ohio State 54.4
the equations for offense/defense ratings, 5 Georgia 53.0
unless we add another equation to guar-
antee a unique solution (this, in fact, is TABLE 5.5: Top 5 Offense
what Massey did in his original work).
To give the offensive and defensive ratings
simple interpretations, the extra equation we 1 Ole Miss 25.9
add forces the sum of the defensive ratings 2 LSU 24.6
to be 0. The top five offensive and defen- 3 Stanford 24.0
sive teams in the 2014 college football reg- 4 Alabama 22.9
ular season are shown to the right. Because 5 Arkansas 21.0
the defensive ratings sum to 0, the average
defensive rating is 0. A team’s offensive rat- TABLE 5.6: Top 5 Defense
ing is therefore the number of points that
team would score against an “average” defense. The defensive rating is the
difference in the number of points that an opponent would score against that
team compared to an “average” defense. The quotes are to bring attention to
the fact that average depends on which teams are included in the ratings. In
this case, all 250 Division 1 football programs were included. By Alabama’s
standards, an average Division 1 school is not very good.

Least Squares Equivalence

Kenneth Massey actually started with a different model for his system.
Suppose that in the first game of the year team A beat team B by 3 points.
We would like to have the ratings’ prediction of a−b match the actual outcome
of 3. However, if in the second game between these teams A won by 2, we have
a problem. We can’t have a − b equal to both 3 and 2.
80 Sports Math

The solution is to acknowledge that the predictions can’t all be perfect,

but our goal can be to minimize the errors. For the two games mentioned
above, the errors would be (a − b) − 3 and (a − b) − 2. For several reasons, it
is common to use the sum of the squares of the errors. For these two games,
that would be (a − b − 3)2 + (a − b − 2)2 . The values of a, b and so on that
minimize the sum of the squares of the errors will be our ratings.
It turns out that the least squares solution for this problem is identical
to the Massey point system described above. Thus, the Massey system makes
sense on the global level (matching net points for the season) and the local
level (giving the most accurate game-by-game predictions).

Wins versus Points

We have two different Massey systems; which is better? Recall that “bet-
ter” depends on the objectives. Let’s take as our main objective the most
accurate prediction of future games (although the eye test of matching public
opinion is also of concern). For convenience, we use “pr” for the point ratings
and “wr” for the win ratings.
The table to the right shows the suc- Win % v.Spread
cess rate of the two rating systems during pr 75.8 54.1
the 2014 college football season. The favorite 15 wr 73.2 55.1
(the team predicted to win by the point spread 74.5 -
spread) won 74.5% of the time. The Massey
points system predicted winners 75.8% of the time, and beat the point spread
54.1% of the time. The results are good if not spectacular, especially for a
system that utilizes such a small amount of information. Note the odd result
that the point ratings did better picking winners, whereas the win ratings did
better against the spread.
It seems reasonable to combine the two Win % v.Spread
ratings. Here is one way to do so. First, put cr 72.3 56.4
the ratings on a similar scale: 15 wr does a mr 74.5 56.4
good job of predicting scores (55.1% against
the spread in 2014). Then find the right proportion of each rating, such as 60%
points and 40% wins. In other words, we want a combination a pr + b (15 wr)
with a + b = 1. Linear regression gives a ≈ .6 as the right choice for the 2014
season. Then cr = .6 pr + 9 wr did a good job of predicting college football
games in 2014. The table shows its record, along with that of mr = 4 pr +
12 wr, the combination that I have used for years.
The statistically inclined should note that the point and win ratings are
highly correlated (ρ = .98), making multiple regression risky. However, pr and
15 wr−pr are only weakly correlated (ρ = −.06), so a regression of the form
pr + c(15 wr−pr) is reasonable.
 Ratings Systems 81

The Colley System

Wesley Colley’s rating system was a mainstay of college football’s BCS sys-
tem. Its origins date back to the mathematician Pierre-Simon Laplace (1749-
1827). When two competitors play for the first time, an unbiased estimate
would give each a probability of 12 of winning. If A beats B in round one, we
might update the probabilities to something like 23 for A. Laplace suggested
1+w
that after w wins in g games, the formula gives a reasonable probability.
2+g
Note that when no games have been played the formula gives the desired
probability of 21 . With successive wins, the formula approaches but never
reaches 1, whereas with successive losses the formula approaches but never
reaches 0. Also, it is common to treat a tie as half of a win and half of a loss.
Objective: A team’s rating is based on Laplace’s probability
1+w
with w modified to include strength of schedule.
2+g
Colley’s inspiration was to rework Laplace’s probability to incorporate
schedule. First, rewrite w as (w − l)/2 + (w + l)/2. Now, if a team played an
average schedule, then the average rating of its opponents would be 12 and the
sum of its opponents’ ratings would be g/2 = (w + l)/2.
1+w
Model: A team with w wins in g games has a rating of with
2+g
w replaced by (w − l)/2 plus the sum of its opponents’ ratings.
To see how this works, let’s return to our four-team league. Team A has
w − l = 5, g = 5, and opponent ratings that sum to 2b + 3d. Then
5
1+ 2 + 2b + 3d
a= .
7
Multiply across by 7 and move the b and d terms to the left to rewrite this
equation as
5 7
7a − 2b − 3d = 1 + =
2 2
which is similar to the Massey win rating equation of 5a − 2b − 3d = 5. The
team’s rating is multiplied by g + 2 instead of g, and the right-hand side is
1 + (w − l)/2 instead of w − l (net wins).
Although they start with very different  7 −2 0 −3 3.5 
models, the Massey and Colley systems end −2 6 −2 0 1 
up with similar equations. The small differ-  0 −2 6 −2 1 

ences in equations make a large difference −3 0 −2 7 −1.5
in their solution. The Colley system has a
unique solution. In this case, we get (rounded to 2 decimal places) a = .76,
b = .57, c = .43, and d = .24.
Notice that the average rating in the above example is .5. You will show
82 Sports Math

in the exercises that this is always the case. An improvement in one team’s
rating is accompanied by equal declines in others.
As formulated, the Colley system uses only schedules and won/lost records.
The similarity in matrix equations of the Massey win and Colley systems gives
us a way to adjust the Colley system to use points. Replace the right-hand
side of the Colley equations with net points, and solve the system. For our
example, we get rounded values of a = 2.32, b = .58, c = −.58, and d = −2.32.
The teams are separated by fewer points in this system than in the Massey
point system, where the gaps between teams were a consistent 2 points.

A Flaky Scaling Problem

There are several properties that it just
seems reasonable for a ranking system to 1 Florida St
possess. Suppose a league plays a complete 2 Alabama
season and you apply a ranking system. 3 Ohio State
Then, in the ultimate deja vu, the season 4 Oregon
gets replayed with exactly the same result. 5 TCU
Instead of team A beating team B by 12
points, we now have team A beating team B TABLE 5.7: Colley Top 5
by 12 points twice. It seems silly to run the
ranking system on this double season, right?
Except that the Colley system changes!
To the right are the Colley ratings for 1 Florida St
the 2014 college football regular season in 2 Alabama
Table 5.7, and the Colley ratings for the 3 Oregon
2014 double season in Table 5.8. Ohio State 4 Ohio State
and Oregon switch positions! This is not an 5 TCU
isolated phenomenon: only 7 of the top 20
teams maintained their original ranks, with TABLE 5.8: Colley Top 5
teams jumping up and down by as many as
3 positions.
How weird is this? Recall that the Colley system is based on the Laplace
formula 1+w2+g . If a team wins 3 out of 5 games, we update their Laplace proba-
bility to 47 . If the team wins 3 of their next 5 games, we update the probability
7
to 12 . This does not seem odd at all. If the team consistently wins 3 out of 5
games, we want our rating to converge to the probability 53 from its starting
point of 12 . You can check that 12 7
is closer to the destination 35 than is 47 .
Perhaps it is not illogical for the ratings to change, even if it is disconcerting
that the rankings would change.
 Ratings Systems 83

The Elo System

Arpad Elo (1903-1992) developed a system to rate chess players. His sys-
tem, slightly modified, is still used to compute the official chess ratings. More
recently, Jeff Sagarin’s adaptation of the Elo system to rate college football
teams was a part of the BCS system.
Objective: A team’s rating is updated after every game based on
its performance in that game, with the change in rating reflecting
the result and the quality of the opponent.
Think about a league in which every team plays on given days (for example,
once a week). After each round of games, the Massey and Colley systems
essentially start from scratch, compiling the schedules and records of each
team for the entire season. By contrast, the Elo system takes the previous
ratings and makes adjustments based on the most recent results. The rating
of a team that just won increases, while the rating of a team that lost decreases.
Model: A team’s rating rnew equals its previous rating rold plus
an adjustment k(s − m), where k is a constant, s is the team’s perfor-
mance in its last game, and m is its predicted performance in that
game.
We will explore each part of this model in turn. First, you need a set of
ratings to adjust once games are played. You could be unbiased and give each
team an equal initial rating, or you could use the best prediction available to
seed the ratings (this could be records from the previous year, or some other
form of preseason rating).
The constant k controls how much the ratings will change based on the
outcome of one game. For chess, the values of k range from 25 for new players
down to 10 for experienced experts. For sports like soccer or tennis, different
values of k can be used to indicate the importance of a match (e.g., 60 for a
World Cup match down to 20 for a friendly). A value of k = 32 has worked
well for the NFL.
The game performance s can indicate winners or point spreads. In chess,
s = 1 for the winner, s = 0 for the loser, and both players get s = .5 for a
pf + 1
draw. For a sport with points, Langville and Meyer suggest s =
pf + pa + 2
for a team that scored pf points and allowed pa points. Note the use of the
Laplace formula, so that 0 < s < 1 and, more importantly for the properties
of the system, the sum of the s values for the two opponents equals 1.
The predicted performance m is intended to quantify the gap in abilities
of the two opponents as essentially a probability that each player will win.
The value suggested by Elo was
1
m=
1 + 10−(ra −rb )/400
for team A entering the game with rating ra against team B rated at rb . This
84 Sports Math

may look intimidating, but Figure 5.1 shows that m increases smoothly from
0 to 1 as the difference in rating goes from −∞ to ∞.

FIGURE 5.1: Elo Performance Graph

An example should help. Person A enters a chess match with rating 2600,
facing person B whose rating is 2200. Based on the ratings, we would have no
trouble picking A to win. For player A, we compute m = 1/(1 + 10−400/400 ) =
1/(1 + 0.1) = 0.91. We would predict A to win 10 out of 11 matches, or 91%.
For player B, we compute m = 1/(1 + 10−−400/400 ) = 1/(1 + 10) = 0.09. We
predict B to win 9% of the time. Note that the probabilities sum to 1.
Now, suppose B wins in an upset. We want B’s rating to increase and A’s
to decrease, but by how much? If A and B were young players, we might want
B’s new rating to be higher than A’s. If they had played often (say, 11 times,
with A winning 10), we would not want B to jump above A. Taking the value
of k to be 10, we compute the new ratings: b = 2200+10(1−.09) = 2209.1 and
a = 2600 + 10(0 − .91) = 2590.9. Notice that B’s rating increased by 9.1 and
A’s decreased by the same amount. This symmetry is built into the system.
This example should make the significance of the choice of k clearer. Since
s − m will always be between 0 and 1, k is the maximum number of points
that a rating can change. For a given sport, its value should be adjusted so
that the ratings are stable (don’t change too much) but responsive (reflect
changes in performance).
Notice also that the 400 in the exponent plays an important role in con-
trolling the rate at which the probabilities approach 0 and 1. If 400 is changed
to 1600 in the above calculation, the probability that A wins drops to 0.64.
A ratings difference of 1600 would be required to reach a win probability of
0.91. This is explored further in exercise 5.17.
 Ratings Systems 85

Strength of Schedule
Each of the systems presented in this chapter incorporates strength of
schedule. This is a critical feature of any rating system. Most people have
a good intuitive understanding of what the concept of strength of schedule
represents. However, a precise determination of how to calculate strength of
schedule turns out to be surprisingly difficult.
Think about teams A and B, each of which plays 10 games. Team A’s
opponents are all good, but not great, teams. Team B plays the three best
teams, three mediocre teams, and four bad teams. Which team has a tougher
schedule? Depending on which system you use to rate teams, team A’s op-
ponents may have a higher average than team B’s opponents. However, there
are other ways to measure strength of schedule.
If both A and B are mediocre to good teams, then A’s schedule is loaded
with maybe win/maybe lose games, while B has 4 sure wins and 3 more easy
games. A reasonable prediction might be 7 wins for B against 5 wins for A.
So A’s schedule looks tougher.
However, what if A and B are both great teams? While A’s games are all
losable, A would be heavily favored in every game; 10-0 and 9-1 are reasonable
predictions. B, however, has three very tough games and 7-3 and 8-2 results
are very possible. Now B’s schedule looks tougher.
Both arguments are valid. If you are trying to pick the top four teams
for a playoff, however, the second argument should feel more convincing. But
here we run into more complications. For the top teams, does it make more
sense to compute an expected record or to compute the probability of going
undefeated?
The bottom line here is that strength of schedule is one of those terms that
sounds simple enough that we don’t necessarily question the assumptions that
go into its calculation. When you see a strength of schedule rating, you should
try to find out how it is computed.

Computing Probabilities
One way to compute strength of schedule is to calculate the probability
that a team of a certain strength could go undefeated. This, of course, requires
that we compute the probabilities that teams of certain strengths will win
particular games. If a team with rating 100 faces a team with rating 94, what
is the probability that the better team wins?
Phrasing the question in this way may prompt you to ask an important
question. How much luck is there in the game? One team is considered to be
86 Sports Math

6 points better, but is that significant? Six goals in soccer or hockey are hard
to score, but six points in football can easily result from a weird bounce of
the ball.

FIGURE 5.2: Predicted Minus Actual Scores, 2014

We will explore this question more fully in Chapter 8, but for now we give
one answer. For the 2014 college football regular season, the Massey ratings
(actually the combination ratings named “mr” above) were used to predict
all Division 1-A games. The histogram in Figure 6.2 shows the frequencies
of values of predicted score minus actual score. The plot looks like a nice
bell curve. The mean is essentially 0, and the standard deviation is about
16 points. That is, start with the Massey prediction and nearly one-third of
the games ended more than 16 points different than the prediction (including
two that were more than 40 points different)! That is a lot of plus/minus. In
that context, a 6-point difference does not seem very large. If the game score
differences are really normally distributed with mean 0 and standard deviation
16, the probability that a 6-point favorite wins is approximately 64.6%. (The
probability that a 16-point favorite wins would compute to be 84%; does that
surprise you?)

Weighty Issues
One objection to the Massey and Colley systems as described here is that
all games are weighted equally. Games that are played early in the season
count as much as games that are played at the end of the season. Thus, a
team that improves during the season will be performing at a higher level
than its rating indicates. In some sports (soccer, golf, tennis, and others), not
all contests have equal importance. Matches in a major tournament should
count more than glorified scrimmages that no one cares about.
Fortunately, both the Massey and Colley systems are easily tweaked to
 Ratings Systems 87

count the most important games extra. In particular, there is no requirement

that the number of games be an integer. You could count an especially im-
portant game (late in the season, or in a major tournament) as 1.5 games, or
an unimportant game as 0.7 games. Winning a 0.7 game would give you 0.7
wins, or winning a 0.7 game by 10 points would give you 7 net points.
What is the best way to weight games? You should not be surprised to
read that “best” depends on your objectives. In a 16-game season, do you
want all games to count equally or not? If not, should game 16 count more
than game 15, or should games 13-16 count the same (and more than games
9-12)? Here are three types of functions you might use for weighting games in
time.
Linear: The weight function is a line. If the last
of n games counts as b games and the first as a
games, then the i-th game counts a + (i − 1)(b −
a)/(n − 1) games. Every game is given a different
weight, with the difference in weights between the
2nd and 3rd games equal to the difference in weights
between the 15th and 16th games.
Step: The weight function is a sequence of stair
steps. The last n1 games count as b1 games, the n2
games before that as b2 , and so on. In the figure
shown, the stairsteps are not equally-sized. You can
weight groups of games as you wish. Here, the last
4 games all carry equal weight.
Logistic: The weight function increases in an S-
shape. To increase from a games to b games, the
b−a
i-th game counts a + for constants c, d,
1 + ced(i−w)
and w that can be adjusted. While all games have
different weights, the last few games are weighted
nearly equally, as are the first few games. There is a
large change in the weights in the middle games.
The proper choice of weights depends on the details of the sport and the
objectives of the ratings system.

Calculus Box: A Recipe for Reduction of Matrices

Technically, this is a linear algebra box to demonstrate a method for find-
ing the reduced form of a matrix. There are numerous calculators and software
packages that will do this for you. As you will see, the method is not compli-
cated, but requires extensive arithmetic; in other words, it is an ideal computer
application.
88 Sports Math

Let’s start with the matrix for the  5 −2 0 −3 22 

Massey points system for the league in Ta- −2 4 −2 0 0 
bles 5.1 and 5.2. The elementary row opera-   0 −2 4 −2 0 

tions we will use are based on equation solv- −3 0 −2 5 −22
ing techniques you have used before. For ex-
ample, if you needed to solve the system of
two equations and two unknowns with x − y = 2 and 3x + 2y = 11, the
technique of elimination can be used. Multiply the first equation by 2 to get
2x − 2y = 4 and then add it to the second equation 3x + 2y = 11. We get
2x − 2y + 3x + 2y = 4 + 11 or 5x = 15, from which we conclude x = 3.
Equivalently, we can eliminate a non-zero element from a row of a matrix by
adding a constant times one row to another row. The algorithm that follows
using the standard matrix notation that the (i,j) entry is the number in the
i-th row and j-th column. For example, in our matrix the (1,5) entry is 22, the
(2,2) entry is 4, the (3,2) entry is −2, and so on.
Row Reduction Algorithm:
1. Start with i=1.
2. Make the (i,i) entry nonzero, swapping rows if necessary.
3. Divide each number in the i-th row by the (i,i) entry.
4. For each j6=i, make the (j,i) entry 0, replacing row j with the
old row j minus the old (j,i) entry times row i.
5. Increase i by 1 and repeat steps 2-4 until finished.
Looking back at our matrix, there is a 5 in the (1,1) entry, so step 3 tells
us to
 divide by 5. The first row is now
1 −2/5 0 −3/5 22/5
The (2,1) entry and (4,1) entry are not yet zero, so we apply step 4 to
each. Multiply the new row 1 by the old (2,1) entry (−2) and subtract it from
the second row. That is,

−2 4 −2 0 0
− −2 4/5 0 6/5 −44/5
= 0 16/5 −2 −6/5 44/5

Our goal of a zero in the first slot is accomplished. We need to do the same
for the fourth row. Take row 4 minus −3 times row 1 (this is the same as row
4 plus 3 times row 1).

−3 0 −2 5 −22
+ 3 −6/5 0 −9/5 66/5
= 0 −6/5 −2 16/5 −44/5

This completes step 4. The updated ma-

 
1 −2/5 0 −3/5 22/5
trix is shown to the right. Moving on in the 0 16/5 −2 −6/5 44/5 
algorithm to i = 2, we implement step 3 by 0 −2
 
4 −2 0 
dividing the second row by 16/5. This is the 0 −6/5 −2 16/5 −44/5
same as multiplying by 5/16.
 Ratings Systems 89

The matrix with its new second row is 1 −2/5 0 −3/5 22/5

shown to the right. We will need to execute 0 1 −5/8 −3/8 11/4 
step 4 three times, to create zeros in the first, 
0 −2 4 −2 0 

third, and fourth rows. Letting the computer 0 −6/5 −2 16/5 −44/5
do this is looking like a good idea. For j=1,
we take the first row plus 2/5 times the second row.

1 −2/5 0 −3/5 22/5

+ 0 2/5 −1/4 −3/20 11/10
= 1 0 −1/4 −3/4 11/2

Notice that we retain the one in the first slot. Since we are working on the
second column (i=2) the desired 0 is in the second column. The next step is
row 3 plus 2 times row 2.

0 −2 4 −2 0
+ 0 2 −5/4 −3/4 11/2
= 0 0 11/4 −11/4 11/2

Finally, we take row 4 plus 6/5 times row 2.

0 −6/5 −2 16/5 −44/5

+ 0 6/5 −3/4 −9/20 33/10
= 0 0 −11/4 11/4 −11/2

This completes step 4. The updated ma-

 
1 0 −1/4 −3/4 11/2
trix is shown to the right. We now move on 0 1 −5/8 −3/8 11/4 
to i = 3. To create a 1 in the (3,3) slot, we
 
0 0 11/4 −11/4 11/2 
implement step 3 by dividing the third row 0 0 −11/4 11/4 −11/2
by 11/4. This is the same as multiplying by
4/11.
The matrix with its new third row is
 
1 0 −1/4 −3/4 11/2
shown to the right. We again have three ze- 0 1 −5/8 −3/8 11/4 
ros to create, two above and one below the
 
0 0 1 −1 2 
(3,3) entry. The work is shown below, start- 0 0 −11/4 11/4 −11/2
ing with row 1 plus 1/4 times row 3.

1 0 −1/4 −3/4 11/2

+ 0 0 1/4 −1/4 1/2
= 1 0 0 −1 6

We next want row 2 plus 5/8 times row 3.

0 1 −5/8 −3/8 11/4

+ 0 0 5/8 −5/8 5/4
= 0 1 0 −1 4
90 Sports Math

Finally, we take row 4 plus 11/4 times row 3.

0 0 −11/4 11/4 −11/2

+ 0 0 11/4 −11/4 11/2
= 0 0 0 0 0

Putting this together gives us the desired reduced matrix.

 
1 0 0 −1 6
0 1 0 −1 4
 
0 0 1 −1 2
0 0 0 0 0

Exercises
In these exercises, T refers to thinking problems, conceptual problems requir-
ing no calculations. C refers to problems requiring calculus, linear algebra
or significant computer calculations. P refers to projects; these are ideas for
further investigation (hints and resources are at the book’s web site).

Team W L PF PA A B C D
A 5 0 74 50 A - 2 1 2
5.1 Suppose that B 2 3 58 53 and B 2 - 2 1 give
C 3 2 52 47 C 1 2 - 2
D 0 5 40 74 D 2 1 2 -
a league’s results and schedule. Set up the matrices needed to find the
(a) Massey win ratings; (b) Massey points ratings; (c) Colley ratings.
 
1 0 0 −1 1.42
0 1 0 −1 0.67
5.2 The reduced matrices for exercise 5.1 round to (a)  0 0 1 −1 0.75;


0 0 0 0 0
   
1 0 0 −1 8.46 1 0 0 0 .775
0 1 0 −1 6.5  0 1 0 0 .475
(b) 
0 0 1 −1 5.29; (c) 0 0 1 0 .525. Find the ratings and
  

0 0 0 0 0 0 0 0 1 .225
rankings for the three systems, and comment on any interesting features. Deter-
mine a scaling factor c such that c times the Massey win ratings are of the same
magnitude as the Massey point ratings. Give one reason why you can know in
advance that, unlike the example in the text, there is no c that will produce an
exact match.
5.3 Suppose that one more game is played in the league of Tables 5.1 and 5.2,
with C beating A by a score of 12-6. Set up the matrices needed to find the
(a) Massey win ratings; (b) Massey points ratings; (c) Colley ratings.
 Ratings Systems 91
 
1 0 0 −1 1.13
0 1 0 −1 0.97
5.4 The reduced matrices for exercise 5.3 round to (a)  0 0 1 −1 0.81;


0 0 0 0 0
   
1 0 0 −1 4.75 1 0 0 0 .67
0 1 0 −1 4.31 0 1 0 0 .57
0 0 1 −1 3.88; (c) 0 0 1 0 .54. Find the ratings and rank-
(b)    

0 0 0 0 0 0 0 0 1 .23
ings for the three systems, and comment on the difference that one game can
make early in a season.
5.5 Set up the (nine) equations for the Massey offensive and defensive ratings for
the league in Tables 5.1 and 5.2. In this case, the ratings are the “least squares
solution” of the equations. (See exercise 5.34 below.)
5.6 The solutions for exercise 5.5 (in order for teams A, B, C, and D) are offensive
ratings of 14.46, 10.93, 10.07, and 7.82 and defensive ratings of −.64, .89, −.25,
and 0. Verify that the defensive ratings sum to 0 and compute the point ratings
as the sum of the offensive and defensive ratings. The point ratings are different
from those reported in the text. Explain the differences.
5.7 For the teams of exercises 5.5 and 5.6, predict the scores of games between
(a) A and B; (b) A and C; (c) A and D; (d) B and C; (e) B and D; (f) C and D.
Note that game scores should be integers.
Team W L A B C D
A 2 0 A - 2 0 0
5.8 Suppose that B 0 2 and B 2 - 0 0 give a league’s re-
C 1 1 C 0 0 - 2
D 1 1 D 0 0 2 -
sults and schedule. Explain why this league’s schedule is not connected. At this
point, is there any basis for comparing, for example, teams A and C? Set up the
matrices needed to find the (a) Massey win ratings; (b) Colley ratings.
 
1 −1 0 0 1
0 0 1 −1 0
5.9 The reduced matrices for exercise 5.8 are (a)  0
;
0 0 0 0
0 0 0 0 0
 
1 0 0 0 3/4
0 1 0 0 1/4
(b) 0 0 1 0 1/2. For part (a), explain why this is not the standard


0 0 0 1 1/2
form for a reduced Massey matrix. Translate the matrix back into equation form
to interpret the results. Explain why these ratings make sense. For part (b), read
off the Colley ratings. Explain why these ratings make sense within the context
of the model for the Colley system.
5.10 In the league of exercise 5.8, if teams A and C play then the schedule
becomes connected. Suppose that C beats A. Set up the matrices needed to find
the (a) Massey win ratings; (b) Colley ratings. (c) What do you expect to happen
to the ratings? Explain.
 
1 0 0 −1 −1
0 1 0 −1 −2
5.11 The reduced matrices for exercise 5.10 are (a)  0 0 1 −1
;
0
0 0 0 0 0
92 Sports Math
 
1 0 0 0 .53
0 1 0 0 .27
(b) 0 0 1 0 .63. Find the ratings and rankings for the two systems


0 0 0 1 .57
and discuss the results. In particular, does it make more sense to you to have
teams C and D tied or to have C above D?
5.12 Look up points scored and allowed per game for the 2014 college football
regular season. How do the top five in Tables 5.5 and 5.6 compare? Explain why
there are differences between the ratings and the actual point totals.
5.13 The Massey point ratings can be tweaked to determine home court advan-
tage. Suppose a league has the following results after a home-road round-robin in
which every team plays each of the other teams twice, once at each team’s home
court. Team A scored 34 points more than its opponents in the three games it
played at home, but scored 10 points less than its opponents in the three games
Team HPD RPD
A 34 -10
it played on the road. B 12 3 Consider each team as being dif-
C -3 -20
D 4 -20
ferent when playing at home and on the road. Thus, there is team AH (team A
playing at home) and a separate team AR (team A playing on the road). Set
up the matrix for the Massey point ratings for the eight teams in this league.
The ratings for the home teams end up being 14.2, 4.3, 1.5, and 4.2 and the
road ratings are 0, 7.6, 0.9, and 0. Discuss any interesting features. In particular,
why is BR rated above BH even though its point differential of 3 is worse than
BH’s point differential of 12? Why are AR and DR rated the same? To compute
each team’s home court advantage, subtract its road rating from its home rating.
Discuss any interesting features.
5.14 Show that the Colley ratings will always average 1/2. (Hints: add the
columns of the Colley matrix and convert this into an equation. Simplify the
equation to conclude that the sum of the ratings equals n/2, where n is the
number of teams in the league.)
5.15 Show that the average Elo rating is constant as long as the sum of the
results s1 + s2 and the sum of the expected results m1 + m2 both equal 1. (Hint:
compute the two changes in ratings resulting from a match.)
5.16 Experienced chess players with ratings 2800 and 2600 play each other. Com-
pute the new ratings if (a) the better player wins; (b) the lesser player wins.
5.17 For chess ratings, show that every 400 points difference indicates that a
player is 10 times better than the opponent. To be precise, if A is ranked 400n
points higher than B for some positive integer n, show that the probability of A
winning is 10n times the probability of B winning.
5.18 If experienced chess players A and B both start with ratings of 2600, A
beats B, and then B beats A, are A and B rated the same? Should they be?
5.19 The USGA golf handicap system uses the average of the best 10 of the last
20 rounds. For a golfer whose last 20 rounds, in order, produce adjusted scores
of 6, 11, 8, 4, 14, 9, 6, 8, 7, 8, 8, 11, 13, 10, 9, 8, 12, 9, 7, and 9 (small is better)
compute the average. What should happen if the golfer’s next score is 7? What
does happen? What happens if the score after that is 16? Comment on this
system.
 Ratings Systems 93

5.20 Elo ratings for the NFL use k = 32 and an exponent of −(ra − rb )/1000. If
A is rated 110 and B is rated 100, update the ratings if (a) A wins; (b) B wins.
5.21 Repeat exercise 5.20 replacing the 1000 in the exponent with (a) 1500; (b)
500. Which value in the exponent is most reasonable? Explain.
5.22 T Two teams finish the season with the same record. They played each
other once. Do you think that the winner should definitely be ranked above the
loser? Give arguments both for and against.
5.23 T Most sports leagues are divided into conferences and/or divisions. Sup-
pose that a team that does not win its conference/division makes the playoffs
and wins the league championship. Does that indicate that the regular season is
meaningless or that the playoffs are highly random? Discuss.
5.24 T If a league wanted to choose the “best” teams (the ones playing the
best at the end of the season) for the playoffs, how could it do so? Describe a
modification of one of the ratings systems in this chapter that could be used.
5.25 T Voters in college sports polls are asked to rank teams from 1-25. Discuss
advantages and disadvantages of allowing the voters to rate the top 25 compared
to the top 1 or top 10.
5.26 T Label each of the following as a rating or ranking problem. (a) set-
ting point spreads for upcoming games; (b) determining a conference champion;
(c) determining seeds for a playoff; (d) the lists in the opening graphic for the
chapter.
5.27 T Compare the models for the Massey, Colley, and Elo systems and state
which one seems to be the most reasonable. Explain.
5.28 T Name several factors that could be important for rating teams but that
are not used by the Massey, Colley, or Elo systems.
5.29 T The four-team 2014 college football playoffs included Alabama, Oregon,
Florida State, and Ohio State. Based on Tables 5.3 and 5.4, was the selection
committee more influenced by wins or points? Comment on whether the com-
mittee was more focused on the “right” teams or the “best” teams.
5.30 T If team B is a heavy underdog but upsets team A, should B’s rating go
up a lot more than A’s rating goes down? Does the Elo system do this? Answer
in terms of both numerical change and percentage change.
5.31 T Explain how strength of schedule is used in each of the Massey, Colley,
and Elo systems.
5.32 T If a league’s schedule and results are doubled (the scenario in “A Flaky
Scaling Problem”), do the Massey ratings change? Explain. Do the Elo ratings
change? Explain.
5.33 T Early in the season team A is outstanding, but several injuries in mid-
season cause A’s quality to drop to mediocre by the end of the season. Should
A’s rating indicate a mediocre team or one that is halfway between outstanding
and mediocre? Should teams that played A early in the season get credit for
playing an outstanding team, or for playing one with A’s rating at the end of
the season? Discuss.
5.34 C Here is some of the linear algebra behind exercise 5.5. The left-hand
sides of your nine equations can be formed into a 9x8 matrix M, with a 9x1
vector b formed by the right-hand sides. Ideally, we would just solve Mx = b
94 Sports Math

for x, but there is no exact solution. We find the best match by multiplying the
equation by MT , the transpose of M. The solution of MT Mx = MT b is our rating
vector reported in exercise 5.5. Compute MT M and MT b and solve the system.
5.35 C Explain in linear algebra terms why every column summing to 0 guar-
antees that the set of row vectors is linearly dependent. Explain in terms of
schedules and results why the set of row vectors is linearly dependent. (Hint: if
you know the schedules and results of all but one team, can you determine the
schedule and results of the remaining team?)
5.36 C Reduce the matrices of exercise 5.1 (a)-(c) to obtain the matrices shown
in exercise 5.2.
5.37 C Assuming a normal distribution of scores with mean 0 and standard
deviation 16 and assuming all games are independent, find the probability that
a team of rating 90 goes undefeated against each schedule. (a) 5 games against
teams of rating 80; (b) games against teams of rating 92, 90, 88, 60, and 60.
5.38 C For the schedules of exercise 5.37, (a) which one has the higher mean
rating? (b) Which one is harder to go undefeated against? (c) Which one is
harder to win at least 3 games against?
5.39 C Pick your own sport and year and rate teams using the Massey point,
Massey win, and Colley systems. An interesting version of this assignment is to
take the schedules and records at the halfway mark of the season and use the
first-half ratings to predict the eventual standings.
5.40 P Develop your own rating system by modifying one of the systems in
this chapter. Be explicit about the reasons for your modifications: what are your
objectives and why are your changes likely to address them?
5.41 P Choose a weighting system and apply it to the rating system and league
of your choice. Explain your choice of weighting system.
5.42 P Research the Keener method or some rating system not covered in this
chapter and apply it to the examples in this chapter or to your favorite sport.
Discuss the apparent advantages and disadvantages of this method.

Further Reading
Amy Langville and Carl Meyer’s Who’s #1? is an excellent resource for
the ratings systems covered here plus several more.
Tim Chartier and Drew Pasteur contributed excellent chapters to Mathe-
matics and Sports edited by Joe Gallian.
An impressive amount of information can be found at Kenneth Massey’s
www.masseyratings.com. There are also good web sites for the Colley, Elo,
and Sagarin systems.
The UMAP Module referenced for the Massey system is number 725 “A
Mathematical Rating System” by Roland Minton. Thomas Jech’s 1983 Ameri-
cam Mathematical Monthly article is what got me started. My “blogging”
about my rating system can be found at my By the Numbers web site.
Chapter 6
Voting Systems

Introduction
Sports fans love to argue. We get espe-
cially engaged (and sometimes enraged) when
the topic is ranking the best players and teams
of all time. We take our Most Valuable Player
(MVP) and Hall of Fame (HOF) voting seri-
ously.
Both MVP and HOF voting will be featured
in this chapter, but the primary focus is on the
mathematics of voting. Whether Pete Rose or
Barry Bonds deserves to be in the Baseball HOF
in Cooperstown is a question for a different book,
perhaps an ethics text. Whether Roger Maris be-
longs in Cooperstown is a question for a differ-
ent chapter, where we look at baseball analytics.
Here, we will analyze the various voting methods
that are used to determine MVPs and HOFs.
A quick scan of sports stories at almost any
time shows the importance of voting in sports. The Heisman Trophy in college
football, the Baseball HOF, the MVPs in baseball and pro football, the NCAA
basketball polls, and countless cell phone surveys all involve voting of one sort
or another.
We will analyze many of the voting methods used in these cases, trying to
determine what biases or injustices might be promoted by the voting rules.
Among the questions we will answer are the following. Which types of voting
systems are used? What are the flaws in these systems? Is there a best voting
system? What do Google and Amazon have to do with voting? Why was
Kenneth Arrow awarded a Nobel Prize for his contributions to voting theory?
Is it better to be seeded ninth or tenth in the NCAA Basketball Tournament?

95
96 Sports Math

How They Vote

We start by reviewing some of the voting methods used in sports.
The Heisman Trophy is the most prestigious award in college football.
The award is decided by a group of approximately one thousand voters, with
each voter listing his or her top three players. Players receive 3 points for
each first-place vote, 2 points for each second-place vote, and 1 point for each
third-place vote. The player with the most points wins.
One of the closest Heisman votes occurred in 2008. Oklahoma’s Sam Brad-
ford won with 300 first-place votes, 315 second-place votes, and 196 third-place
votes for 300(3) + 315(2) + 196(1) = 1726 points. Colt McCoy of Texas was
second with 266 first-place votes, 288 second-place votes, and 230 third-place
votes for 1604 points, and Tim Tebow was third with 309 first-place votes,
207 second-place votes, and 234 third-place votes for 1575 points. Is there
anything that bothers you about this? For some, the fact that Tebow had
more first-place votes than Bradford means that Tebow should have won. A
counterargument is that 811 voters had Bradford in the top three compared
to 750 voters for Tebow.
The MVP voting in professional football will not have this issue. Voter
Barry Wilner said, “We don’t want an MVP who doesn’t get the most choices,
so we use a better system that guarantees the player with a majority of first-
place votes wins.” The NFL MVP is decided by 50 voters who each vote for
one player. The player with the most votes wins. In most years, the winner
does indeed have a majority of votes, but in 2005 Shaun Alexander won with
19 of the 50 votes (38%).
We need two definitions. A majority of votes is more than 50% of the
votes. For the 50 votes in the NFL’s system, a majority is 26 or more votes.
(25 votes, exactly 50%, does not constitute a majority; only one person can
have a majority of the votes.) Shaun Alexander did not have a majority of
votes in 2005. Alexander, however, did have a plurality of votes: a plurality
of votes is the most first-place votes received by any candidate. Most elections
in the United States use the plurality system. When there are more than two
strong candidates, the plurality winner could have a small number of votes. As
we have seen, people do not always accurately distinguish between majority
and plurality.
Baseball and basketball do not share the NFL’s objection to the point
system (or its confusion of majority and plurality). The major league baseball
(MLB) MVP system asks each voter to list ten players in order, with players
receiving 14 points for each first-place vote, 9 points for each second-place vote,
8 points for each third place vote, and so on down to 1 point for each tenth-
place vote. Voters for the Cy Young Award for best pitcher list 5 players, who
receive 7, 4, 3, 2, or 1 point for votes. Point systems like this are susceptible
to the 2008 Heisman situation of the plurality winner not being the overall
 Voting Systems 97

winner. In 1966, Roberto Clemente edged Sandy Koufax for National League
MVP in spite of receiving only 8 first-place votes to Koufax’s 9.
The NBA uses a point system for its MVP with the top five for each voter
receiving 10, 7, 5, 3, or 1 point for votes. In 1990, Magic Johnson was elected
MVP despite receiving only 27 first-place votes, compared to Charles Barkley’s
38 first-place votes. Magic received 38 second-place votes to Barkley’s 15 to
more than make up the point difference.
The Baseball Hall of Fame uses a different type of voting system. From
a list of eligible players, each of the approximately 500 voters may vote for
up to ten players whom the voter deems worthy of inclusion. Any number of
votes from zero to ten is allowed. To be elected, a player must be listed on at
least 75% of all submitted ballots.
For the selection of the host city of an Olympic Games, cities are elimi-
nated until one city receives a majority of first-place votes. In voting for the
1994 Winter Olympics, 84 voters each voted for one city. In the first round,
Lillehammer received 25 votes, Anchorage 23, Ostersund 19, and Sofia 17.
Since the plurality winner, Lillehammer, did not receive a majority of votes
(25 out of 84 is less than 30%), a second round of voting was needed. The
last-place city from the previous round, in this case Sofia, was dropped and
another vote was taken. This time Ostersund received 33 votes (39%), Lille-
hammer 30, and Anchorage 22. Anchorage was dropped and on the third vote
Lillehammer received 45 votes (54%, a majority!) to win. If it strikes you as
odd that Ostersund went from third to first to second, you might wonder if
this is a good voting system.
While “majority rules” may seem like the obvious approach to any election,
we are about to see why we need alternatives.

Condorcet’s Intransitive Attitude

The transitive property is an example of an idea that we often take for
granted, but which we assume to be true at our own peril. The real numbers
have an ordering, so that we can always rank them, with the transitive
property that if a < b and b < c, then a < c. This means that once we know
that 2 < 3 and 3 < 5, we can immediately jump to the conclusion that 2 < 5.
In sports, if we believe that Dallas is better than New York and New York
is better than Washington, then we surely believe that Dallas is better than
Washington. This belief is necessary if we are going to rank teams.
Trivial and boring? Try this. Early in the season, Dallas has defeated
New York and New York has defeated Washington. Does this guarantee that
Dallas will defeat Washington? Of course not. Upsets happen, teams have
bad days, balls bounce funny, and referees make calls. Sports results are not
98 Sports Math

transitive. This, in fact, is part of the fun of following sports: the possibility
of a surprising, illogical result is always there.
Transitivity is a basic assumption underlying ranking systems. When we
list our top 5, we mean that number 1 is better than number 2, and number 1
is better than number 3, and so on. However, we recognize that sports results
are not transitive. Voting, it turns out, is decidedly not transitive.
Suppose that a three-person committee needs to rank three teams A, B,
and C. One person ranks the teams in the order A-B-C, the second person
ranks them in the order B-C-A, and the last person ranks them in the order
C-A-B. Each team is tied with one first-place vote. Believing in the majority-
rules principle, we look at two teams at a time. Given a choice between only A
and B, two out of three committee members (the first and third) prefer A over
B. Also, two out of three committee members prefer B over C. The committee
prefers A over B, and prefers B over C; do we even need to check out A versus
C? Actually, we do, because two out of three committee members prefer C
over A!
In voting situations with three or
more candidates, the transitive law does
not hold in the sense that the majority
of voters can prefer candidate A over B,
B over C, and C over A. The problem
is not one of quirky personalities or vot-
ers changing minds or being illogical. The
transitive law is simply not true for vot-
ing situations of this type.
The intransitivity of pairwise voting
has been known for hundreds of years,
going back to the first analytical study
of voting. In 1785, the Marquis de Con-
dorcet published a theory of voting, a
logical outgrowth of the rational ideal-
ism spawned by the American Revolution
and its influence in France. Condorcet
proposed the two-at-a-time process de- FIGURE 6.1: The Author and
scribed above, and stated that the candi- Condorcet in Paris
date who wins all head-to-head matchups
should be the overall winner. The problem, as Condorcet himself discovered,
is that there is not always such a winner. In our three-team example above,
no team wins all of its head-to-head battles; we say that A, B, and C form a
Condorcet cycle.
For Condorcet, the search for a perfect voting system hit a dead end. For
us, the lack of a transitive property for group voting means, for starters, that
voting methods are different from ranking systems.
 Voting Systems 99

Preference Lists, Voting Systems, and Chaos

We look at situations where a group of voters is trying to rank some
number m of candidates with m > 2. Each voter has a preference list of
candidates; i.e., a ranking of candidates. The preference list for each voter
obeys the transitive property; an individual who prefers A over B and B over
C also prefers A over C. Based on this collection of preference lists, we want
to identify the winner of the election.
We have already defined the plurality method. The candidate who re-
ceives the most first-place votes wins the election; we call this candidate the
plurality winner. The NFL MVP is decided by the plurality method. In
this method, as with the others to follow, there is the possibility that a tie
will need to be broken. For simplicity, we make the assumption that ties are
broken in some logical manner (which we will not address specifically).
The Condorcet winner of an election is a candidate, if one exists, who
wins every head-to-head matchup with other candidates. The problem, as
we have seen, is that in some cases there is no Condorcet winner, due to a
Condorcet cycle. Of course, a necessary property for a useful voting method
is that it determines a winner in every election, so this is a big problem. This
is also an issue for the majority method, in which a candidate receiving a
majority (more than 50%) of the votes is the winner.
A popular voting method is the Borda count. Each candidate receives
m − 1 points for each first-place vote, m − 2 points for each second-place vote,
and so on down to 0 points for each last-place vote. The Borda winner is
the candidate with the most total points. The Heisman Trophy, the MLB
MVP and the NBA MVP are versions of the Borda count. Voters for these
awards are not asked to rank all candidates, however, and the points awarded
vary. The Borda count was invented by Condorcet’s contemporary and rival,
Jean Charles de Borda, and later re-invented by Charles Dodgson (aka Lewis
Carroll of Alice’s Adventures in Wonderland fame).
The Olympic Games method of selecting host cities is an example of a
plurality with elimination method. In this method, successive votes are
taken until a candidate receives a majority of votes; this candidate is declared
the winner. If a vote does not result in a majority decision, the candidate with
the fewest votes is eliminated and a new vote is tabulated. For the Olympics,
separate votes are taken, allowing voters to change their minds. In instant
runoff voting, each voter submits a preference list and all rounds of voting
are conducted from this list. For example, a voter with preference list A-B-C-
D would cast a vote for A; if A is eliminated, that voter’s vote would go to B.
If B is also eliminated, the vote next goes to C, and so on.
The approval voting method allows voters to vote for any and all candi-
dates of which they approve. That is, for each candidate the voter says “yes”
or “no” and the candidate with the most yes-votes is the approval win-
100 Sports Math

ner. The Baseball Hall of Fame voting is similar to approval voting, although
each HOF voter is limited to no more than ten votes. Further, all candidates
who receive more than 75% of the possible votes are elected, so there can be
multiple winners.

4 6 9 11
A A B C
Example 6.1 For the preference lists given, find the fol-
B C A B
C B C A
lowing (if it exists): (a) majority winner, (b) plurality winner, (c) Condorcet
winner, (d) Borda winner, (e) plurality with elimination winner.
Solution. The table indicates that 4 voters rank the candidates in the order
A-B-C, 6 voters rank the candidates in the order A-C-B, and so on. So 4+6=10
voters cast first-place votes for A, 9 cast first-place votes for B, and 11 cast
first-place votes for C.
(a) With 30 voters, a majority is at least 16 votes, so there is not a majority
winner. (b) C has the most first-place votes (11) so C is the plurality winner.
(c) Matching A and B head-to-head, the 4 voters in the first column and the
6 voters in the second column prefer A over B, while the remaining 9+11 =
20 voters prefer B over A. B beats A, 20-10. Matching B and C, the 4 voters
in the first column and the 9 voters in the third column prefer B over C, but
the other 6+11 voters prefer C; C beats B, 17-13. Matching A and C, A gets
the voters in the first three columns and wins 19-11. Each candidate wins one
and loses one head-to-head matchup, so there is no Condorcet winner.
(d) For the Borda count, note that A gets 4+6=10 first-place votes, 9 second-
place votes, and 11 third-place votes. Then compute
A : 2x10 + 1x9 + 0x11 = 29 points
B : 2x9 + 1x(4+11) + 0x6 = 33 points
C : 2x11 + 1x6 + 0x(4+9) = 28 points
so B is the Borda winner. (e) For plurality with elimination, start with first-
place votes of 10 (A), 9 (B), and 11 (C). Nobody has a majority, so we drop
the lowest vote-getter, which is B. With B eliminated, A now gets the 9 votes
from B’s supporters and beats C 19-11. A wins plurality with elimination.

Example 6.1 should be disturbing. Depending on the voting method used,

any of the three candidates can win! Of course, this example was specially
constructed to work this way. In many elections, there is a clear-cut winner
and the voting method used does not matter. In close elections with three or
more candidates, however, you should be aware that the winner might depend
on the voting method used.
 Voting Systems 101

Fairness and the Arrow of Impossibility

In a close election, different voting methods can produce different results.
This leads us to investigate which method is the best and which methods
should never be used. Taking this negative angle, we want to reject any method
that behaves in an illogical manner. The fairness criteria make precise some
of the ways that a voting method could misbehave.
We want voting methods to satisfy monotonicity: if candidate A would
win using a set of preference lists and the only changes that are made to the
preference lists move A higher in the list(s), then A should still win. Think
back to Example 6.1(e). With the voter preference lists as given, A wins using
plurality with elimination. Now, suppose that right before the vote occurs
A makes an announcement that sways some of the voters in A’s favor (no
jokes about the Qatar World Cup selection, please). Before looking at the
details, let’s summarize: A was going to win the election anyway, and then
gains more support. Logically, A should still win, right? This is what we mean
by monotonicity.
Split the fourth column of 11 voters in Example 6.1 into 8 who maintain
the same order and 3 who move A to the top (making their preference A-C-B).
Now A has 13 first-place votes, B has 9 votes and C has 8. We eliminate C,
and then match A and B head-to-head. B gets 9+8 votes and beats A, 17-13,
to win the election. What happened? Three voters improved their opinion
of A and this cost A the election! Technically, we say that plurality with
elimination violates the monotonicity fairness criterion. More informally, we
can cross that method off the list of good methods.
Another important fairness criterion goes by the unwieldy name of in-
dependence of irrelevant alternatives (IIA): if a set of preference lists
is changed in a way that leaves the relative ordering of candidates A and B
the same, then the voting method should not change its ordering of A and
B. In other words, once the voting populace has decided that it prefers A to
B, additions, deletions, or re-orderings of other (irrelevant) candidates should
not reverse the decision.
A famous violation of IIA occurred at the 1997 European Men’s Figure
Skating Championships. At the time, the system used to rank competitors was
complicated (see exercise 6.19), but involved two rounds of skating. With one
skater left in the final round, the official standings were Vyacheslav Zagorod-
niuk in first, Alexei Urmanov second, Ilya Kulik third, and Philippe Candeloro
fourth. The last skater ended up finishing sixth, so logically the top four should
remain the top four, right? However, the order changed completely: Urmanov
moved up to first, Candeloro was second, Zagorodniuk third, and Kulik fourth.
Zagorodniuk fell from first to third without skating, and with only the sixth-
place skater doing anything! This final result was definitely dependent on the
irrelevant alternative of the last, sixth-place skater.
102 Sports Math

Unfortunately, both plurality and the Borda count can violate IIA. In
Example 6.1(b), C wins plurality over A and B, but if in the third column A
and B are reversed (which should be irrelevant since both are already ranked
above C in this list) then C loses and A wins. In Example 6.1(d), B wins the
Borda count, but if A and C are flipped in the second column then C wins.
We now have reasons to throw out all five of our voting methods: majority
and Condorcet winner do not always produce winners, and the other three
violate fairness criteria. The hope of finding a perfect voting method was
skewered in 1948 by Kenneth Arrow, who won a Nobel Prize for this and other
work. Arrow’s Impossibility Theorem states that among voting methods
that rationally use voter preference lists to rank three or more candidates,
the only method that does not violate at least one of the two fairness criteria
(monotonicity and IIA) is a dictatorship (that is, a situation where a single
voter makes all decisions). In short, there is no perfect voting method based
on preference lists. Note that approval voting does not use full preference lists
(if a voter approves of candidates A and B, we cannot tell whether or not A
is preferred over B) and so is not judged by Arrow’s Impossibility Theorem.
As we will see, approval voting has its issues.
There are further properties we might hope that a voting method would
satisfy. For example, when a majority winner or Condorcet winner exists, we
want our voting method to identify that candidate as the winner. Approval
voting violates both of these criteria (as does the Borda count; plurality vio-
lates the Condorcet criterion). Each voting method has its own set of eccentric
behaviors that make it appear unfit for usage.
As discouraging as that may seem, mathematically speaking it just means
that the research continues. If there is no perfect method, there could still be
a best method, one which is least likely to violate fairness criteria or one which
is optimal by some other measure. The search continues in the next section.

Positional Voting Systems

Mathematical results have been proved for voting systems that can be
framed as positional voting systems. The concept is easy to understand
with the Borda count. In Example 6.1, there are three candidates and the
Borda count assigns 2, 1, and 0 points to votes. We can represent this method
as a vector [2, 1, 0]. If we want to emphasize first-place votes by assigning them
5 points, our system is represented by [5, 1, 0]. If we want to use plurality, the
system is described by [1, 0, 0], since only a first-place vote counts. Approval
voting would be described by [1, 1, 0] if all voters approved of exactly two
candidates, but since this is not always the case approval voting is not a
positional voting system.
 Voting Systems 103

Positional voting systems include three of the voting methods discussed

above. Plurality with elimination and approval voting do not fit this system.
Mathematicians have done extensive work on positional voting systems,
and have proved a result that is suggestive for sports voting. The result relates
to positional voting systems with full ranking of candidates (each preference
list ranks all candidates). Sports situations where a voter submits a top ten
or top twenty are not full rankings, since most teams are not included in the
lists.
However, in the case of full rankings the Borda count is the positional
voting system that maximizes the likelihood that the Condorcet winner is
ranked first. Thus, changing a [2, 1, 0] Borda count to a [5, 1, 0] system with a
bonus for first place votes makes violations of the Condorcet winner criterion
more likely.
This does not prove that the unequal weights for MVP votes are subopti-
mal, since MVP votes are not based on full rankings. However, it does suggest
that excessive tinkering with the points for first place could be counterpro-
ductive.

A Return to Sports Voting

With some mathematical terminology and results in hand, let’s return to
our sports examples for some evaluations.
The Heisman Trophy uses a (partial ranking) Borda count [3, 2, 1]. There
are problems with the Borda count. As Tim Tebow found out, the plurality
winner also has no guarantee of being the Borda winner; even a majority
winner can lose the Borda count. The Borda count also violates IIA.
On the plus side, the Borda count allows voters to give support to a variety
of candidates. Further, suppose that there are two top candidates A and B
for the Heisman. You favor A. In one case, you find it difficult to distinguish
between A and B; in another case, you think that B is unworthy of the award.
With plurality, you vote for A. With the Borda count, you can give B second-
place points or worse, so your opinion of B counts.
The NFL MVP is a plurality vote. In close elections, the plurality system
is unusually prone to violating head-to-head preferences of the voters, as we
see in the simulation results that follow. The main positives of the plurality
system are its ease of implementation, the lack of knowledge required of voters
(you only need to pick out one candidate; perhaps for an important award this
should not be considered a positive), and its familiarity.
Baseball and basketball MVPs are decided by (partial ranking) Borda
counts with extra points for first place votes. We see in Figures 6.2 and 6.3
that the extra points make little difference in the ability of the system to
agree with head-to-head preferences. Philosophically, this type of system is
104 Sports Math

a compromise between the broad-based support of a Borda count and the

familiarity of the plurality system. Mathematically, it can be viewed as a
weighted sum of the Borda count and plurality systems, with the plurality
component polluting the better properties of the Borda count.
The Olympics selection process uses plurality with elimination. This sys-
tem violates IIA and monotonicity, with the monotonicity issue making the
process seem almost random for close elections. On the plus side, a Condorcet
winner (if one exists) will always be elected and elimination eventually boils
down to a majority-rules decision.
The baseball HOF uses a modified approval voting method, with each
voter limited to ten approvals and all candidates with at least 75% of the
votes elected. Approval voting does not directly violate Arrow’s Impossibility
Theorem, but the HOF modification does violate IIA (see exercise 6.26). The
method used allows for unequal numbers of candidates to be elected in differ-
ent years, giving the process a feeling of a knowledgable thumbs up or thumbs
down on each candidate.

Simulations
Other than the Olympics selection committee, sports voters do not use a
pure version of any of the basic voting methods. While this is done for sound
practical reasons (we should not really care whether a voter ranks Sammy
Watkins 23rd or 24th in MVP voting), it does mean that the mathematical
beauty of Arrow’s Impossibility Theorem does not directly apply. If there
are no theoretical results to evaluate our voting methods, we can still collect
evidence from simulations. Because most simulations include some element(s)
of randomness, simulation results are rarely conclusive. Nonetheless, seeing
the voting methods in action can give us important information.
The first round of simulations applies to 50 players who have a “true”
ability. The abilities are not bunched: the best player has a value of 50, the
second best 49, and so on. The voters have a fuzzy evaluation of the players:
each voter rates each player as v + e where v is the true value of the player,
and e is a random error in judgment. So, a given voter might rate the first
player as 50−3.4 = 46.6 and the second player as 49+0.3 = 49.3 and conclude
that the second player is better. The simulation takes these flawed ratings and
orders them into preference lists for 60 voters. The preference lists are then
run through three voting methods: plurality, a [20, 19, 18, ..., 2, 1] partial
ranking Borda count for the top twenty, a [25, 19, 18, ..., 2, 1] count that gives
5 bonus points for first-place votes, and approval voting (scores above 45 - on
average, the top five - are approved). This is done 1000 times and the results
compiled.
To see which voting methods follow the “will of the people,” we look at
 Voting Systems 105

how often the voting methods agree with the opinions of the voters (whether
they were “right” or not). If a majority of voters prefer A to B, we check
whether the voting method ranks A over B. If it does, the voting method
“wins” that matchup. For all four methods, there are “ties” where both A
and B receive 0 points. This is especially true for the plurality method, since
the 50th best player is unlikely to receive a first-place vote. Ties are ignored,
and we compute a winning percentage for the cases where the voting method
has an opinion.

FIGURE 6.2: Borda Count, Approval, Plurality

Figure 6.2 shows the results for a variety of average voter evaluation errors.
The higher the error, the less certain the voters are about the true quality of
the players, and therefore the more random their votes are. Only three curves
are distinguishable, because the winning percentages for the two Borda counts
agreed to three decimal places. For small errors, the ordering of players is clear
to most voters, and all voting methods perform well. When there is substantial
disagreement among the voters as to who is best, plurality does a noticeably
worse job of matching voter sentiment than the Borda count and approval
voting.
The results change little when the first-place bonus points increase to 20.
When the threshold for approval voting is lowered from 45 to 40, the method
does better for large voter evaluation errors, essentially matching the Borda
count winning percentages.When the number of voters is increased to 300, the
winning percentages of all three methods increase.
A second round of simulations explores what happens when two contro-
versial candidates are in the pool. The controversy may be over the meaning
of “valuable” (e.g., a player on a losing team) or over some social issue (e.g.,
a steroid user). For each controversial candidate, about one-third of the vot-
ers start with a base value that puts the candidate in first place. The other
two-thirds have a starting value in tenth place (losing team) and twentieth
(steroids).
Plurality is clearly inferior here. Although its winning percentage increases
as voter errors increase, this is not a ringing endorsement for a voting method.
106 Sports Math

(“It’s better when nobody knows what they’re voting on!”) You can see the
effect of giving bonus points for first place in this simulation, with the bonus
Borda distinctly below the regular Borda count.

FIGURE 6.3: Two Controversial Candidates

It is important to repeat that simulations do not prove anything. However,

plurality does not look good here. We sometimes unthinkingly act like this is
the only way to vote, but most voting methods experts consider plurality one
of the worst voting methods available. By association, the bonus Borda also
looks inferior. The voting methods research community has a lively debate
over whether the Borda count or approval voting is better. The simulations
give an edge to the Borda count if most voters only approve of a small number
of candidates, with the two methods tied for more generous approval voters.

Range Voting
The old figure skating voting system gave us an egregious violation of IIA,
when the leaderboard at a major competition was suddenly scrambled. To
avoid such embarrassments, the scoring system was completely revised. The
new system is a point system, where nine judges give the skaters grades, the
highest and lowest are discarded, and the average of the seven remaining scores
is retained. Similar systems are used in other sports, including gymnastics and
diving.
This is called a trimmed mean, and the voting system is similar to a
system called range voting. We use range voting regularly to provide feed-
back on purchases at such places as Amazon. Range voting is not based on
preference lists. Voters assign each candidate a rating (e.g., five stars) and
may give the same rating to many candidates. Because the outcome is not
based on preference lists, Arrow’s Theorem does not apply to range voting.
In fact, range voting does not violate monotonicity (moving a candidate up
 Voting Systems 107

gives the candidate more points, which can only help) and does not violate
IIA (the points assigned to two candidates do not change if other candidates
are added, deleted, or modified).
Even though it passes the criteria of Arrow’s Theorem, range voting can
violate a majority decision if the minority ranks the candidate low enough.
This indicates that range voting is susceptible to insincere or strategic voting.
Suppose you want player A to win the MVP, and it is clear that player B is
A’s main competition. Even though you respect player B, you might decide to
improve A’s chances by giving A the maximum vote and B the minimum vote.
If you also give other candidates the minimum score, what you have done is
equivalent to plurality voting, giving one candidate one vote. Since plurality
does not rate highly as a voting method, strategic voting could reduce range
voting to an inferior product.
You might expect that figure skating, which has a history of manipulated
votes, is subject to this flaw. However, it seems that reviews of judges’ records
has been enough to keep most judges honest.

FIGURE 6.4: Range vs Borda

The results from adding range voting to the simulation of Figure 6.2 are
shown in Figure 6.4. To simulate range voting, each voter’s candidate ratings
is converted to an integer from 1 to 5. (Using the actual voter rating is more
accurate, but it would be unrealistic to ask real voters to provide ratings with
3 significant digits.) As it is, range voting is better than the Borda count, and
hence better than approval voting and plurality, at identifying head-to-head
voter preferences.
A sophisticated simulation published in Gaming the Vote took into account
a variety of possibilities, including insincere voting and voter ignorance of some
candidates. In this case, the criterion for best voting method was to maximize
the overall “happiness” of the voting populace. Range voting was clearly better
than approval voting, Borda count, and the other methods discussed here. Of
these, plurality came in last.
108 Sports Math

PageRank and MVPassing

A search engine gives you an ordered list of results. You expect that the
first results listed will be the most useful, but how does the search engine
decide which results to present first? In this section, we discuss the original
formulation of Google’s PageRank for determining the importance of a web
page. The importance of the page is combined with the relevance of the page
to the keywords in the search to produce the order of the search results.
A voting analogy help us understand how PageRank works. View web
pages as shareholders. Suppose that page A has 10 shares of “stock” and links
to five other pages. Then A is voting 2 votes apiece to those 5 pages. In turn,
A gets its votes when other pages link to A. The more links A has, and the
more shares each of those pages has, the higher A will rank.

Imagine a “web” of four pages. Page A links to pages B and C; if A has a

shares, then B and C each receive 12 a votes. If page B links to A and D, each
receives 12 b votes. If page C only links to A, then A receives c votes. If page
D links to all three pages, each receives 13 d votes. In all, page A has received
1 1 1 1
2 b + c + 3 d votes, so a = 2 b + c + 3 d, and we have similar equations for b, c,
and d.
Does this look familiar? It should remind you of equations for the Massey
and Colley systems in Chapter 5. PageRank can be thought of as a matrix-
based rating system. Another view of PageRank is developed in exercise 6.23.

Example 6.2 For the four-page web described above, find PageRank values
that sum to 1.
Solution. The equations for the ratings a, b, c, and d are given by
a = 12 b + c + 31 d
b = 12 a + 13 d
c = 12 a + 13 d
d = 21 b
which can be summarized in the matrix
 Voting Systems 109
 
−1 1/2 1 1/3 0
1/2 −1 0 1/3 0
 .
1/2 0 −1 1/3 0
0 1/2 0 −1 0
You can verify that this system of equations has an infinity of solutions. This
is why we add the requirement that the ratings sum to 1; that is, we add the
equation
 a + b + c + d = 1 toget the matrix
−1 1/2 1 1/3 0
1/2 −1 0 1/3 0
 
1/2
 0 −1 1/3 0 ;
 0 1/2 0 −1 0
1 1 1 1 1
andthen reduce this matrix
 to get
1 0 0 0 10/25
0 1 0 0 6/25 
 
0 0 1 0 6/25 ;
 
0 0 0 1 3/25 
0 0 0 0 0
so a = 10/25, b = c = 6/25, and d = 3/25. A is the most important page,
B and C are tied for second, and D is last. If you think through the linking
structure, this should make sense. All pages link to A, and only B links to D.
B and C are both linked to by A and D.

There are interesting applications of PageRank to sports. Using PageRank

to rank teams, as in Chapter 5, is not one of them. For web pages, a page
votes for another page if it links to that page. For teams, we could say that a
team votes for another team if it loses to that team. The issue of what to do
with an undefeated team is explored in exercise 6.23. A larger problem occurs
with a team that loses only once. This is a very good team and should be
highly rated. The team that beats it gets all of its voting shares. In the case
of a large upset, this gives the underdog team far too much credit.
PageRank has been applied with good
effect to passing networks in sports such
as soccer and field hockey. Soccer teams
have different passing tendencies. Barcelona
is known for incessant short passes, while
other teams are known for long balls. We
can count the number of passes made be-
tween players, and this may give us informa-
tion about tendencies. The data may show
that a team favors the left side, but that is likely to be known to the players.
A harder task might be to identify the most important link in the passing
structure (the MVPasser), analogous to the most important page on the web.
Suppose our team has four players. During the course of the game, player
A passes to player B 6 times, player C 12 times, and player D 2 times. This is
110 Sports Math

shown in the graph above, along with the remainder of the passing data. The
arrows of the graph (we call them directed edges) are sized according to how
many passes are made, so we can see quickly that A and D do not interact
much.
The
 matrix for the ratings a, b, c, and
 d is given by
−1 8/30 12/40 4/20 0
 6/20
 −1 14/40 5/20 0 
12/20 15/30
 −1 11/20 0 .
 2/20 7/30 14/40 −1 0
1 1 1 1 1
The first column shows that 6 out of A’s 20 passes are directed to B, 12 out
of 20 to C, and 2 out of 20 to D. Reducing the matrix gives us the ratings
a = .209, b = .237, c = .354, and d = .200. This tells us that player C is the
MVPasser. In this case, the ratings have direct meaning. Over the course of
an infinitely long game, C has control of the ball 35.4% of the time, B has
the ball 23.7% of the time, A 20.9%, and D 20%. The percentages are not
measured in clock time; a better way of phrasing the result is that 35.4% of
the passes go to player C.

Seeding of Tournaments
The seeding of tournaments is designed to separate the best players or
teams so that they do not face each other early in the tournament. In the
NCAA basketball tournament, all teams are seeded. Do the seedings matter?
In the spirit of our voting method analysis, we want to ask whether any
paradoxes occur in which it is better to have a worse seed.
At first glance, the answer appears to be “no.” Suppose that the tourna-
ment has 64 teams in four regions. In each region, teams are ranked 1-16. In
the first round, the best team (seed 1) plays the worst team (seed 16), the
second-best team (seed 2) plays the second-worst team (seed 15), and so on.
Figure 6.5 shows the frequency with which a given seed won its first round
game in the years 1979-2014. With the exception of seeds 5 and 12, the per-
centage is decreasing in an orderly linear fashion. Seeds 5 and 12 play each
other, and it is a well-known part of “March madness” that the 12-seeds al-
ways provide an upset victory or three. In the first round, the better the seed
the better the result.
Ignoring the 5-12 blip, we can fit a line y = 1.0627 − .0662x to the data
for the probability that seed x wins its first-round game. Our question about
seeding is not fully answered, however. In the second round games, we expect
to have seeds 1 and 8, 2 and 7, 3 and 6, and 4 and 5 play. However, suppose
that both the 9-seed and 10-seed win in the first round. Then 1 plays 9 and
2 plays 10; the 10-seed has an easier game! To see how these matchups play
 Voting Systems 111

out over the course of the tournament, we can use simulations. To do so, we
need probabilities for teams winning games in each round.

FIGURE 6.5: First Round Wins by Seed

The expected second-round games (1-8, 2-7, 3-6, 4-5) follow a similar trend
to Figure 6.5. The exception is the 6-seed beating the 3-seed more often than
expected. The line y = .896 − .083x is a good fit. For example, 2-seeds beat
7-seeds with probability .730 while 1-seeds beat 8-seeds with probability .813.
If the 10-seed upsets the 7-seed in the first round, a good approximation for
the 2-10 second round game is to boost the 2-seed’s probability one level, from
.730 to .813. The logic is that 1-8 and 2-10 games are similar in the degree of
mismatch. Implicit in the logic is a re-evaluation of the quality of the teams.
A 15-seed who has reached the third round has proven to be underrated, and
so is given a better chance to win. Figure 6.6 shows the results of a million
simulated tournaments.

FIGURE 6.6: Predicted Wins by Seed, Rounds 1-4

112 Sports Math

Beyond round 1, it is not true that better Seed Actual Pred.

seeds get better results. Figure 6.6 shows that 1 39.6 35.4
more 10-seeds, 11-seeds, 12-seeds, and 13-seeds 2 21.5 24.1
win second round games than do 9-seeds. This is 3 11.1 13.2
because 9-seeds have to play the 1-seeds, whereas 4 9.7 8.3
10-seeds play 2-seeds or, sometimes, 15-seeds. In 5 4.8 5.4
the third round, 6-seeds win as often as 5-seeds. 6 4.2 4.4
Again, the explanation is that it is bad news to 7 1.4 3.0
play a 1-seed. In round 3, 5-seeds almost surely 8 4.2 1.5
play the 1-seed. Winners in round 4 (the winner 9 1.4 0.7
of the region) are typically the top three seeds, 10 0 1.1
but seeds 8, 10, 11, and 12 have nearly equal like- 11 2.1 1.1
lihoods of advancing to the Final Four. To the 12 0 0.9
right is our reality check: the actual breakdown
of regional winners for 1979-2014 next to the FIGURE 6.7: Percentage
predictions from the simulations. The biggest in Final Four by Seed
tournament surprises include the three 11-seeds
(VCU, George Mason, and LSU) who reached the Final Four. The simulation
predicts three 11- or 12-seeds out of the 144 regions in 36 years, and that is
exactly what has happened! Yes, lumping seeds together is cheating, but the
predictions are reasonably accurate. Of importance for us is the conclusion
that 10-, 11- and 12-seeds all have a better chance of reaching the Final Four
than 9-seeds.

Probability Box: Put Some Error Bars on Those Things

It may have occurred to you that the above simulation was unnecessary.
Once we have the individual probabilities (e.g., a 3-seed beats a 15-seed with
probability 0.79) it is possible to directly compute the probabilities of reaching
different rounds. The computed probabilities and the Final Four probabilities
shown above agree to the third decimal.
However, simulation results should be accompanied by error bars. In the
NCAA simulation, 35.4% of the 1-seeds reached the Final Four. Will a new
simulation produce a different number? Could it match the actual value of
39.6%? Having a sense of the variation in data allows us to answer such ques-
tions. Ideally, we can compute the theoretical standard deviation; a sample
calculation follows. In the absence of an underlying theory, we can repeat the
simulation numerous times and estimate the standard deviation that way.
As an example, we can track the 1-seeds in the first two rounds. In round
one, the 1-seeds win with probability 0.9965. In probability terms, we model
the games as Bernoulli trials with p = 0.9965; in each game, the 1-seed wins
with this probability. For a simulation of one million trials (n), the standard
 Voting Systems 113
q
p(1−p)
deviation of the proportion of wins equals σ = n = 0.000059. Less
than 5% of the simulations should produce a winning proportion more than
2σ away from p, outside of the range 0.9965 ± .000118. So, the simulated first
round should be very accurate.
To make it through two rounds, the 1-seed must win in the first round
and then beat either the 8-seed or the 9-seed. We multiply probabilities to
get a combined probability of 0.9965 p2 for a second round win, where p2
is the probability that the 1-seed wins its second round game. There are
two possibilities: the 1-seed beats the 8-seed or the 9-seed (whichever won
in the first round). We need the probability that #1 beats #8 (0.813) in
round two times the probability that #8 beat #9 in round one (0.5331),
plus the probability that #1 beats #9 (0.896) times the probability that
#9 beat #8 (0.4669). The probability that a 1-seed wins in round two is
0.9965(0.813 · 0.5331 + 0.896 · 0.4669) = 0.8488.
The calculation of the standard deviation gets messy. With the notation
p1 = 0.9965 and σ1 = 0.000059 for the first round mean and standard devia-
tion, respectively, and p2 and σ2 for the corresponding
p second round values, the
standard deviation for second round wins is p21 σ22 + p22 σ12 + σ12 σ22 . We have
p2 = 0.813 · 0.5331 + 0.896 · 0.4669 = 0.852 and can compute σ2 = 0.000654.
The second round standard deviation is 0.00056. About two-thirds of the
simulations should produce a 1-seed second-round probability in the range
0.8488 ± 0.00056. Our simulation value of 0.8484 is in the lower end of that
region.
The bottom line is that the error bars for Figure 6.6 would be invisible,
because the standard deviations are quite small for simulations of one million
tournaments.

Exercises
In these exercises, T refers to thinking problems, conceptual problems requir-
ing no calculations. C refers to problems requiring significant proof-writing
ability. P refers to projects; these are ideas for further investigation (hints
and resources are at the book’s web site).

6.1 (a) Suppose that in the 2008 Heisman voting points were only given for first
and second place votes, with 3 points for a first place vote and 1 point for a
second place vote. Would the results have changed? (b) Assigning x points for
first place votes, 2 for second, and 1 for third, how large would x have to have
been for Tebow to win the 2008 Heisman? (c) If the Heisman voting were changed
to a HOF-like system where voters vote for three players, who would have won
the 2008 Heisman? (d) Are there any years in which this system would have
changed the winner?
114 Sports Math

6.2 Determine whether the Heisman voting procedure obeys majority rules,
meaning that a player who receives a majority of first place votes always wins.
6.3 For the 2005 NFL MVP voting, Shaun Alexander received 19 of the 50 votes,
followed by Peyton Manning with 13 and Tom Brady with 10. If the voting
system used was to receive 2 points for a first place vote and 1 point for a second
place vote, give a hypothetical circumstance in which the winner could have been
(a) Manning, (b) Brady.
6.4 In the 1999 American League MVP voting, Pudge Rodriguez received 7 first
place votes and 252 points to edge Pedro Martinez, who received 8 first place
votes and 239 points. If both players received 2 7th-place votes, 2 6th-place votes,
2 5th-place votes, 3 4th-place votes, and 5 3rd place votes (this is hypothetical),
how many 2nd-place votes did each player receive? In this situation, how many
of the 28 voters left Martinez off their ballots? Why might a voter do that?
6.5 Explain how you know that at least one voter changed his or her preference
list in the 1994 Olympics vote. Recall that the votes were taken at different times.
Describe a scenario in which it would be reasonable to change your mind.
6.6 Suppose that in voting for the 1994 Olympics site, 25 voters had the prefer-
ence list L/A/O/S, 17 voters had the preference list A/S/L/O, 6 voters had the
preference list A/O/S/L, 19 voters had the preference list O/A/L/S, 14 voters
had the preference list S/O/A/L, and 3 voters had the preference list S/L/A/O.
(These totals are made up.) Determine the winner using plurality with elimina-
tion. If a third- or fourth-place vote represents “disapprove” which city has the
(a) highest (b) lowest percentage of “disapprove” votes?
6.7 In the scenario of exercise 6.6, suppose that at each stage the city with
the most last-place votes is eliminated. Show that Anchorage would be chosen.
Discuss which elimination rule (most last-place votes or fewest first-place votes)
makes more sense.
6.8 In an election with three candidates, 8 voters have the preference list A/B/C,
5 voters have B/C/A, 3 voters have C/A/B, and 3 voters have C/B/A. Find the
winner using (a) plurality; (b) Borda count; (c) plurality with elimination. (d)
Is there a Condorcet winner? (e) If each voter approves of his or her first two
candidates, who wins approval voting?
6.9 In an election with four candidates, 7 voters have the preference list
A/D/B/C, 5 voters have C/A/B/D, 4 voters have C/B/D/A, and 3 voters have
D/A/B/C. Find the winner using (a) plurality; (b) Borda count. (c) Is there a
Condorcet winner? (d) A Condorcet loser is a candidate who loses every head-
to-head matchup. Show that the plurality winner is a Condorcet loser. Comment
on why many feel that plurality is a poor voting method.
6.10 In an election with four candidates, 10 voters have the preference list
A/B/C/D, 8 voters have C/B/A/D, 6 voters have C/D/B/A, and 5 voters have
B/C/D/A. Find the winner using (a) plurality; (b) plurality with elimination;
(c) Borda count. (d) Find the Condorcet winner. (e) Find the Borda count win-
ner if D drops out, leaving three candidates. Which fairness criterion has been
violated?
6.11 For each method, give an example showing that it violates the Condorcet
criterion: (a) plurality; (b) Borda count; (c) approval voting; (d) range voting.
6.12 Suppose teams are rated by some method and every game conforms exactly
to the ratings (i.e., if team A is rated 5 points better than B, then A beats B
 Voting Systems 115

by 5 points). Explain why a Condorcet cycle is not possible. In other words,

Condorcet cycles are caused by deviations from teams’ ratings.
6.13 Verify that (a) if A and B are reversed in the third column of example
6.1(b), C loses plurality; (b) if A and C are reversed in the second column of
example 6.1(b), B loses the Borda count. (c) Explain the significance of parts
(a) and (b) in terms of the fairness criteria.
6.14 The Borda count can be used to rank all candidates; in Example 6.1, the
ranking would be B (33 points), then A (29 points), then C (28 points). Explain
how to use (a) plurality and (b) plurality with elimination to rank all candidates.
(c) Suppose you have these rankings (no ties). If all voters reverse their preference
lists (e.g., change A/B/C to C/B/A), will the rankings be reversed? Determine
yes or no for each of the three voting systems, and comment on whether you
think a good ranking system should reverse the order.
6.15 Show that the positional voting systems [4, 3, 2, 1] and [3, 2, 1, 0] are
equivalent, meaning that they will always produce the same rankings as each
other. Show that the positional voting systems [6, 4, 2, 0] and [3, 2, 1, 0] are
equivalent. Show that [6, 3, 1] and [3, 2, 1] are not equivalent by finding a set of
preference lists in which the systems produce different winners.
6.16 Determine whether the positional voting systems are equivalent or not. (a)
[4, 3, 2, 1] and [7, 5, 3, 1]; (b) [4, 3, 2, 1] and [10, 7, 4, 1]; (c) [4, 3, 2, 1] and [8,
4, 2, 1]; (d) [8, 4, 2, 1] and [27, 9, 3, 1].
6.17 Find an example in which plurality with elimination violates IIA.
6.18 Which of the voting methods (plurality, Borda count, approval, range) vio-
late the monotonicity condition? Explain.
6.19 This problem illustrates the IIA violation in the 1997 European Men’s Fig-
ure Skating Championship. After skating, each skater receives a numerical rating
from each of the seven judges that is converted to an ordinal ranking from that
judge. Suppose that skater A’s ordinals from the seven judges are 1, 1, 1, 3, 2,
2, 4; skater B’s ordinals are 2, 3, 2, 1, 1, 4, 2; skater C’s ordinals are 3, 2, 3, 2, 3,
1, 1. That is, the first judge ranked the skaters in the order A/B/C, the second
judge ranked them in the order A/C/B, and so on. Find the winner using (a)
plurality; (b) Borda count (with five candidates). The skating system was to find
the median ordinal for each skater. (c) Show that each skater has median 2. The
first tie-breaker was to count the number of voters who gave the median mark
or better. (d) Show that skater C finishes third. The second tie-breaker is to add
up the ordinals for the voters who gave the median mark or better. (e) Show
that skater A wins. Now, suppose that there is one more skater to compete. This
skater receives ordinals of 1, 4, 4, 4, 4, 2, 2. (f) Update the ordinals for skaters A,
B, and C. (g) Show that the last skater finishes fourth, and determine the new
top three (you will need a third tie-breaker, adding together all of the ordinals).
(h) Explain why this is a violation of IIA, and discuss how the fans who watched
the event would react to a shuffling of the order.
6.20 Give a positional voting representation for each voting system. (Assume
that there are six candidates) (a) NFL MVP voting; (b) NBA MVP voting; (c)
MLB Cy Young voting.
6.21 In an election with three candidates, 8 voters have the preference list
A/B/C, 5 voters have B/C/A, and 4 voters have C/B/A. Under the positional
voting system [x,2,1], find all values of x such that A wins under the positional
116 Sports Math

voting system [x, 2, 1]. Find all values of x such that (a) A wins; (b) B wins; (c)
C wins.
6.22 (a) Explain all four equations in Example 6.2. (b) Verify that the system of
equations has an infinite number of solutions.
6.23 This exercise gives an alternative explanation of the PageRank equations.
Imagine a bored person online who randomly clicks on links. (a) For the network
in Example 6.2, site A can be linked to from sites B, C, and D. If the amount
of time spent at the sites is named a, b, c, and d, respectively, explain why
a = 21 b + c + 31 d. Interpret the solutions to Example 6.2 with this interpretation.
(b) If a site has no links out, this interpretation falls apart. For the network of
Example 6.2, suppose that site D links to site E, which has no links out. Write
out the matrix for this network. (c) Now, imagine that the bored person types
in a random url to end up at A, B, C, D, or E. Compare the resulting matrix to
one where E links to all other pages. (d) Finally, suppose that 85% of the time
the bored person clicks on a link and the other 15% of the time starts over with
a preference for site A half the time, and B and C a quarter of the time. Write
out the matrix-vector equation for this situation. This is close to the original
PageRank system used by Google.
6.24 Using the linear equations from the text, find the probability of each upset
in the NCAA basketball tournament. (a) #12 beats #5; (b) #9 beats #8; (c)
#15 beats #2; (d) #8 beats #1; (e) #6 beats #3.
6.25 T Pete Rose is not in the Baseball Hall of Fame due to his suspension
for betting on baseball games. Barry Bonds is not in the Hall of Fame due to
his involvement with steroids. Roger Maris had two MVP years but then had
a string of injuries that cut his career short. Discuss your opinions on whether
scandals of various types should disqualify someone from the Hall of Fame, and
what the proper balance of historic performance and longevity should be.
6.26 T Explain why Baseball Hall of Fame voting cannot violate the IIA cri-
terion if each voter has an unlimited number of votes. However, suppose that a
day before the election player A has enough votes for election and B does not.
Then two other players are in the news, one in a positive way and the other in
a negative way. Give a scenario in which these irrelevant players can cause B to
be elected and A not.
6.27 T In each of the following leagues, find a Condorcet cycle: NFL, NBA,
MLB, college basketball, college football.
6.28 T Suppose there are three basketball teams. Team A gives team B fits
with its pressing defense, but team B gives team C problems with its outside
shooting, and team C plays well against team A by breaking the press. Discuss
how a Condorcet cycle could occur, and whether it makes sense to try to rank
these teams.
6.29 T For each method, propose a reasonable way to break ties. Hint: you
may need to separately consider the cases where the number of voters are even
or odd. (a) plurality; (b) Borda count.
6.30 T If there is a Condorcet cycle, is it possible for there to be a Condorcet
winner? Briefly explain.
6.31 T Label each as True or False and briefly explain why. (a) If there is a
Condorcet winner, then there is a majority winner. (b) If there is a majority
 Voting Systems 117

winner, then there is a Condorcet winner. (c) If there is a Condorcet winner,

that candidate will win the Borda count. (d) If there is a majority winner, that
candidate will win the Borda count.
6.32 T If A and B are known to be the best of five candidates for an award,
and you want A to win, how would you order your preference list to give A the
best chance to win a Borda count? If approval voting is used, how would you
vote to give A the best chance? If range voting is used (with a “star” system of
1, 2, 3, or 4) how would you vote to give A the best chance? Discuss the extent
to which these voting systems are subject to strategic voting of this type.
6.33 T Read about slime mold in Ellenberg’s How Not to Be Wrong and explain
how slime mold violates IIA.
6.34 T In an episode of Brain Games, movie-goers showed a strong preference
for a small tub of popcorn over a large tub. When a medium size was introduced
(at a price close to the price of the large), customers chose the large tub over
the small tub. Explain why this is a violation of IIA.
6.35 T Kenneth Arrow used the word “rational” to refer to the transitive prop-
erty. Do you think that it is irrational to have a system that does not obey
transitivity?
6.36 T Explain why the existence of Condorcet cycles proves that “majority
rules” is not transitive.
6.37 T The concept of IIA is that a voting system’s ranking of A and B should
not be reversed based on the actions of some other candidate C. However, this
means that the voting method should ignore the difference between ranking A
first and B a close second compared to A first and B tenth. Do you think that
IIA is a fairness condition that a voting system must obey?
6.38 T Suppose there are 121 voters in an election. (a) How many votes are
needed to be a majority winner? (b) In the extreme, what is the fewest votes
that a plurality winner could receive? (c) How many votes are needed to win
with a 2/3 majority?
6.39 T Discuss the extent to which you think each of the following should be
a factor in MVP voting. (a) playing on a winning team; (b) having the best
statistics; (c) playing the full season; (d) being a good citizen.
6.40 T (a) In Figure 6.2, explain why the winning percentages decrease as vote
uncertainty increases. (b) Explain why the winning percentages increase as the
number of voters increases.
6.41 T (a) For range voting, give an example where one insincere or corrupt
voter could change the outcome. (b) Explain why the trimmed mean option
of dropping the highest and lowest scores is popular. (c) When evaluating the
judges, is it a bad sign if a judge’s score is often thrown out? never thrown out?
Discuss how you might evaluate the judges.
6.42 T Discuss the difference between evaluating voting methods by whether
they violate fairness conditions, versus evaluating voting methods by how
“happy” the voters are with the decisions.
6.43 T In the commentary following Figure 6.7, explain why it is “cheating” to
group together the 11- and 12-seeds.
6.44 T Discuss the advantages and disadvantages of being a 9-seed compared
to an 11-seed in the NCAA basketball tournament.
118 Sports Math

6.45 T In a round-robin tournament, each pair of teams plays exactly once.

Suppose that two teams have the same number of wins. Show that there is a
Condorcet cycle.
6.46 T In a round-robin tournament, player A wins a medal if for every other
player B, either A beat B or A beat a player who beat B. (a) Show that at least
one player wins a medal, and (b) find a circumstance in which very player wins
a medal.
6.47 C In the 1989-90 NBA MVP voting, 92 voters awarded 27 first-place votes
to Magic Johnson, 38 to Charles Barkley, and 21 to Michael Jordan. Johnson won
with 636 points. Suppose that Barkley only received first-place and third-place
votes. How many voters must have left Barkley off the ballot completely?
6.48 C Suppose that the voters have a single issue on which they will base their
votes. In this case, some possible preference lists cannot occur (one of Arrow’s
assumptions is that all lists are possible). Think of a candidate’s position on the
issue as being marked on a football field, ranging from left end zone to right end
zone. Each voter’s stance is also represented by a yard-line, and the preference
list is determined by how close each candidate is to that yard-line. If candidate
A is at the left 10 yard-line, B is at the left 40 yard-line, C is at the right 30
yard-line, and you are at the 50 yard-line, your preference list will be B/C/A
since B is the closest to you, followed by C. Explain why the preference lists
A/C/B and C/A/B cannot occur. In this situation, show that violations of IIA
cannot occur.
6.49 C Suppose that there are only two candidates and an odd number of
voters. Explain why there will be a majority winner, and show that the majority
winner will also win plurality and the Borda count. That is, majority rules works
when there are two candidates. Show that approval voting and range voting will
not necessarily elect the majority winner, however.
6.50 C Six dice are numbered in unusual ways. Die A has all 3s; die B has four
4s and two 0s; die C has three 5s and three 1s; die D has two 6s and four 2s.
Show that the dice violate transitivity, in that A loses to B, which loses to C,
which loses to D, which loses to A.
6.51 C For range voting on a scale of 1 to 5, if a player currently has a range
average of 4.6 and you think the average should be 4.0, can rating the player 1
accomplish your goal? Determine the number of votes for which this could be
true.
6.52 C A small network has 5 sites. A links to B and D; B links to A, C, and
E; C links to B and D; D links to A, B, and E; E links to A and C. Write out
the basic rating matrix (as in Example 6.2) for this network and solve it.
6.53 C A team has five players. During the course of a game, A passes to B 14
times, to C 22 times, to D 18 times, and to E 6 times. B passes to A 10 times,
to C 8 times, to D 12 times, and to E 14 times. C passes to A 4 times, to B 8
times, to D 3 times, and to E 5 times. D passes to A 16 times, to B 6 times,
to C 22 times, and to E 14 times. E passes to A 13 times, to B 12 times, to C
11 times, and to D 10 times. Set up the PageRank matrix for the passer ratings
and reduce the matrix.
6.54 C In Analyzing Wimbledon, Klaasen and Magnus analyze the success of
different seeds not by number (as we did in this chapter) but by expected round
 Voting Systems 119

reached. With 128 players, there are seven rounds. The 1-seed should win 7
matches, the 2-seed should win 6, both the 3- and 4-seeds should win 5, seeds
5-8 should win 4, and so on. Discuss how this compares to computing the log
base 2 of the seedings and rounding up.
6.55 P Develop a program which will take two teams A and B and find the
shortest link of teams such that A beat C beat D beat B. What percentage of
teams can be linked this way? Identify two teams to be in the same group if
there is a link from A to B and a link from B to A. How many different groups
are there?
6.56 P Run the simulation discussed in the section. For each of 1000 voters,
generate the voter’s rating of player 1 (50 plus a sample from a normal distribu-
tion with mean 0 and standard deviation 1) and player 2 (49 plus noise) and so
on, compute the rank and hence the points assigned to that player by that voter,
and add up the points for each player to determine the voting method’s ranking
of the players. For each pair of players, determine which player was preferred
by a majority of voters; if the voting method ranks that player higher, give the
voting method a “win.” If there is no majority winner, do not count the matchup
as a win or a loss. Do this for a variety of voting methods, initial values of the
players, and standard deviations.
6.57 P Run the PageRank system on the sports league of your choice. Compare
the rankings to those obtained from other systems and critique the PageRank
system.

Further Reading
Amy Langville and Carl Meyer’s Who’s #1? is an excellent resource for
the ratings systems covered here plus several more.
The Heisman vote totals are from the Heisman web site:
http://heisman.com/roster.aspx?rp id=74&path=football accessed 4/9/15.
The Barry Wilner quote is from http://profootballtalk.nbcsports.com/
2011/12/28/ap-explains-voting-process-for-nfl-awards/ accessed 4/9/15.
NFL MVP voting history is from http://www.coldhardfootballfacts.com/
content/voting-history-why-peyton-manning-should-run-away-with-nfl-mvp -
award/20626/ accessed 4/9/15.
The NBA MVP voting history can be found at http://www.basketball-
reference.com/awards/ accessed 4/9/15.
The MLB MVP and Cy Young voting history can be found at
http://www.baseball-reference.com/awards/ accessed 4/9/15.
There are several good books about voting methods. Kenneth Arrow’s
Social Choice and Individual Values is an early summary of Arrow’s work.
Jeff Suzuki’s Constitutional Calculus is excellent. Stories about Lewis Carroll
and voting are in William Poundstone’s Gaming the Vote. A journal article of
120 Sports Math

note is in the October 2009 American Mathematical Monthly by Daugherty,

Eustus, Minton, and Orrison.
Google’s description of PageRank with helpful links can be found at
http://google.about.com/od/searchengineoptimization/a/pagerankexplain.htm
accessed 5/27/15.
The June 2014 issue of the Journal of Quantitative Analysis in Sports has
an article that adapts the PageRank algorithm to an effective rating system.
See “An Oracle method to predict NFL games” by Balreira, Miuli, and Tegt-
meyer.
The first issue in 2015 (volume 11, number 1) of the Journal of Quan-
titative Analysis in Sports has several excellent articles about March Mad-
ness. An analysis of the 5-12 matchup can be found at FiveThirtyEight:
http://fivethirtyeight.com/tag/no-12-seeds/ accessed 4/9/15. Peter Keating’s
Giant Killers talk at the 2014 Sloan Sports Analytics Conference is interest-
ing.
Chapter 7
Saber- and Other Metrics

Introduction
In the movie Moneyball the character Pe-
ter Brand, played by Jonah Hill, makes the
case for the use of statistical analysis in base-
ball, “Using stats the way we read them, we’ll
find value in players that nobody else can see.
People are overlooked for a variety of biased
reasons and perceived flaws: age, appearance,
personality. Bill James and mathematics cut
straight through that.”
The one-sentence sound-bite review of
Moneyball is that it describes the conflict be-
tween academic stat-head analysts and old-
school baseball insiders. Such a review is
FIGURE 7.1: Bill James
overly simplistic, but the basic conflict is real,
and is a recurring theme of modern society. As Ian Ayres puts it in his book
SuperCrunchers, “We are in a historic moment of horse-versus-locomotive
competition, where intuitive and experiential expertise is losing out time and
again to number crunching.” However, unless you’re selling books or movies
you should not believe that this is an us-versus-them battle.
If you carefully read Peter Brand’s speech, you can identify the strengths
and weaknesses of the use of statistics in sports. A baseball player’s playing
statistics do not care if the player looks overweight or has an odd technique.
The analyst can “cut straight through” those biases. However, the numbers
may also overlook the odd technique that is likely to lead to injury, or the
temper that makes the player uncoachable. Smart users of analytics (includ-
ing the Oakland A’s of Billy Beane, who now have a larger scouting budget
than they did before Moneyball ) draw on all information to make decisions.
Analytics should not be ignored or blindly followed.
The word “analytics” will appear frequently in the remaining four chapters.
For some, analytics only refers to the products of sophisticated statistical
analysis. In this book, analytics will refer to the identification of patterns in
data through any statistical analysis.

121
122 Sports Math

In this chapter, we take a first look at sabermetrics (baseball analytics)

and analytics in other sports. We will answer the following questions. How
do points scored and allowed relate to wins? What makes a statistic useful?
What are some of the useful statistics that have been developed recently? How
can defense be measured? How can the effect of a distinctive home field be
measured? Is it possible to apply the new statistics to players in the past?

The Pythagorean Cult

Pythagoras (c.571-496 B.C.) was a Greek mathematician and philosopher.
He and his followers are credited with numerous mathematical discoveries and
advances in music, astronomy, and other fields of inquiry. The Pythagoreans
were a secretive bunch, so we know distressingly little of what they actually
did. The theorem that bears Pythagoras’ name (the relationship a2 + b2 = c2
holds in any right triangle of side lengths a, b, and c) was not original to the
Pythagoreans. Its connection to baseball remained hidden for centuries.
Bill James (1949- ) is for many people the father of modern sports analyt-
ics. He did his original work as a night watchman at a pork-and-beans factory.
There are two relevant facts here: he was a baseball outsider, and he was ded-
icated enough to take a job that gave him time to copy and analyze a season’s
worth of box scores by hand. He self-published results in what became a highly
influential series of books, the Bill James Baseball Abstract series (1977-1988).
Because of his outsider status, his ideas and their evolution were very publicly
aired and allowed a generation of sports fans to develop ideas along with him.
Among his first ideas was what he called the Pythagorean Method.
The philosophy behind the Pythagorean Method is crucial. And simple.
The most important statistic for a baseball team is wins. It follows that ev-
erything else we do statistically should relate to the ability to win games. A
statistic should always relate to wins in a measurable way. Backing up one
step, the way you win games is by scoring more runs than the opposition. If we
can relate runs scored to wins, then any statistic that can be related to runs
can be related to wins. The first step is to find an equation that relates runs
and wins. What Bill James found is this: if a team scores RS runs and allows
RA runs, then its winning percentage (more precisely, winning proportion)
will be approximately
RS2
WP = .
RS2 + RA2
I assume that James chose the name Pythagorean Method because of the
sums of squares in the denominator.
 Saber- and Other Metrics 123

W L RS RA
Example 7.1 SF 88 74 665 614
KC 89 73 651 624
For the participants in the 2014 World Series, compute the expected wins
using the Pythagorean Method and compare to the actual wins.
Solution. The table indicates that San Francisco scored 665 runs and allowed
614. The Pythagorean Method predicts a winning proportion of 6652 /(6652 +
6142 ) = 0.5398. Multiplied by 162 games, the Giants are predicted to win
87.45 games, very close to the actual total of 88 wins. For Kansas City, the
Pythagorean Method predicts a winning proportion of 6512 /(6512 + 6242 ) =
0.5212. Multiplied by 162 games, the Royals are predicted to win 84.43 games,
less than the actual 89 wins.

The exponent 2 is used in Example 7.1, but other exponents a in the

RSa
general formula WP = are equally valid theoretically. Which
RSa + RAa
exponent is best? In the absence of theory, the way to answer the question
is to try several exponents and see which one makes the best predictions. By
“best predictions” we mean the smallest differences between actual wins and
predicted wins. A common way of computing a total error is to add together
the squares of the individual team errors.

FIGURE 7.2: The Best Exponent, MLB 2014

Example 7.2 Make a table of errors in Pythagorean Method predictions for

the 30 teams in 2014. Compute the sum of the squares of the errors. Then
repeat this for exponents a = 1.4, a = 1.5, ..., a = 2.3 and find the exponent
that minimizes the sum of the squares of the errors.
Solution. If the error is computed as actual wins minus predicted wins, the
errors in Example 7.1 are 88 − 87.45 = 0.55 and 89 − 84.43 = 4.57. The sum of
the squares of all of the errors with the exponent 2 equals 468.9. This can be
computed in Excel, for example. Changing the exponent in Excel is easy and
quick. The first few lines of a table of errors is shown below. Data are from
baseball-reference.com. All teams played 162 games.
124 Sports Math

W RS RA PY error square
ARI 64 615 742 65.97 -1.97 3.88
ATL 79 573 597 77.68 1.32 1.75
BAL 96 705 593 94.88 1.12 1.26
BOS 71 634 715 71.31 -0.31 0.09
CHC 73 614 707 69.65 3.35 11.21
... .. ... ... ... ... ...
The total error is the sum of all of the entries in the last column shown.
Figure 7.2 shows the sums of squares of errors for the other exponents. To one
decimal, the minimum is 445.3 corresponding to the exponent 1.8.

Theoretically, it can be shown that under certain assumptions the winning

RSa
proportion is WP = for some exponent a. The exponent a = 1.82
RS + RAa
a
is often cited as the best for baseball; for 2014, a = 1.81 is actually slightly
better. Although the scale in Figure 7.2 has been magnified, notice that the
exponent a = 2 works reasonably well for baseball. Exponents for other sports
are found in the exercises.
Pythagorean Method predictions are listed at baseball-reference.com and
other sites. The error is commonly labeled as “luck.” The label implies that
there is no rhyme or reason to the errors. Kansas City won 4 more games than
predicted in 2014; is this because they have a repeatable skill for winning close
games (such as great relief pitching) or did they just have good luck?
Given that Mariano Rivera was the most dominant closer in baseball his-
tory, perhaps the New York Yankees are a good test case. Starting in 2013 and
working backwards, the Yankees’ “luck” value (actual wins minus predicted
wins) was +6 in 2013, 0 in 2012, −4 in 2011, −2 in 2010, +8 in 2009, +2
in 2008, −3 in 2007. There is not an obvious trend there. Three are positive,
three are negative, and one is zero. However, there is a reason I stopped where
I did. The values for 2006 back to 1998 are +2, +5, +12, +5, +4, +6, +2, +2,
and +6. They averaged nearly 5 wins more than predicted for 9 years! This
does not look like luck.
Most teams follow a random-looking up/down pattern of errors, so the
phrase “luck” is appropriate. As with most analytics, however, be aware that
there are exceptions to the rule. The physics of sports deals with events that
must always occur. Analytics uncovers interesting patterns in events that have
happened. The patterns may or may not continue to hold in the future.
Stanley Rothman has provided an interesting alternative to the
Pythagorean Method. He asked a simple question; why does it have to
be so complicated? Many patterns are linear or approximately linear, per-
haps winning percentage is one of them. He proposes the equation WP =
0.000683(RS − RA) + 0.5. The sum of the square errors for 2014 is 407.4, less
than the Pythagorean sum of square errors of 445.3. The average absolute
error of the linear model is also smaller, at 2.82 compared to 2.89. Other than
 Saber- and Other Metrics 125

the constant 0.000683 being ugly-looking, this model is an improvement on

the Pythagorean Method!

When Good Statistics Go Bad

We have established a strong connection between runs in baseball and
wins. Similar connections can be made between points and wins in sports like
basketball and football, or goals and wins in soccer and hockey. The next step
is to relate individual statistics to runs (or points or goals). Since most of the
early work was done in baseball, we develop this idea with baseball statistics.
Some basic definitions may help. Batting average is computed as hits
divided by at bats. Three decimal places are traditional, and an average of
.406 is pronounced “four oh six” and is an average that is “over four hundred.”
The number of at bats does not count plate appearances in which the batter
walks, is hit by a pitch, or is credited with a sacrifice. On-base percentage
(OBP), by contrast, equals hits plus walks plus hit by pitches divided by total
plate appearances. For both statistics, a batter who reaches base due to a
defensive error is debited with an out.
Batting average is a statistic that dates back to the beginnings of the game.
Generations of fans have learned that a batting average of .300 is excellent and
batters near the “Mendoza line” of .200 are not long for the major leagues.
Nevertheless, batting average as a measure of batting performance was one of
the first ideas that Bill James challenged. If batting average does not have a
strong connection to runs scored, then its worth as a statistic is dubious. A
scatter plot is the graph of a set of points representing two variables. In the
scatter plot to the left in Figure 7.3, a point (x, y) consists of a team’s batting
average in 2004 as x and the number of runs the team scored in 2004 as y.

FIGURE 7.3: BA vs OBP, 2004

In both plots, there is a clear trend that points to the right are higher.
That is, teams with better batting averages (or better on-base percentage)
scored more runs. This is not surprising, and indicates that both statistics are
126 Sports Math

reasonable measures of run-scoring ability, and hence winning ability. Which

statistic is better? We want the best relationship possible to runs scored. Try
an example: predict the runs scored for a team with batting average .270. In
the scatter plot, the points above the .270 label show run values from 700 to
850; this is not a very specific prediction. We would like a smaller range of
values, with the points more tightly grouped vertically. The closer the points
come to a straight line, the better.
Visually, it may or may not be clear that on-base percentage gives tighter
predictions. There are ways to quantify this (and avoid any optical illusions
from the graph). Both regression and correlation will be discussed in later
sections. Linear regression is easy enough to visualize in Figure 7.3 as a line
through the center of the data. This line can be thought of as predicting
values of runs scored, so we can use the sum of the squares of errors (SSE)
to evaluate the predictions. Predictions from batting average have SSE equal
to 71539, while predictions from on-base percentage have SSE equal to 47159.
We get much better predictions of runs scored (and hence wins) from on-
base percentage! Correlation indicates how strong a linear relation is; in this
case, the larger the correlation, the stronger the relationship. The correlation
between batting average and runs scored is 0.80, while the correlation between
on-base percentage and runs scored is 0.88.
For these reasons, on-base percentage became known as a better measure of
batting performance than batting average. On-base percentage was a hallmark
of the original Moneyball movement, as players who could draw walks became
more valued and batters were taught to take more pitches (and, in the process,
tire out the opposing pitcher). An even better batting metric is OPS, short for
On-base Plus Slugging (slugging percentage equals total bases divided by at
bats; total bases, in turn, counts 4 for a home run, 3 for a triple, and so on).
It is considered better because it has a stronger relationship to runs scored.
The scatter plot in Figure 7.4 (from 2014 data) is clearly tighter than either
of those in Figure 7.3.

FIGURE 7.4: Runs vs OPS, 2014

The new and better statistics are empirical, meaning that they are data-
driven. An important implication of this is that evaluations may change when
the data change. Compare Figure 7.3 with its 2014 analog in Figure 7.5. Check
 Saber- and Other Metrics 127

the vertical scale of runs scored to see how much scoring decreased in a decade.
In these plots, it is hard to tell any difference between batting average and
on-base percentage. In fact, the SSE values are nearly identical (33518 for
batting average versus 33458 for on-base percentage) and the correlations are
nearly identical (both at 0.797 to three decimals). Compare these to the OPS
statistic, which has SSE equal to 16487 and a correlation to runs scored of
0.91.

FIGURE 7.5: BA vs OBP, 2014

In 2014, then, OBP holds no clear predictive advantage over batting av-
erage. The new truths of one generation may become the outdated dogma of
another generation.

Rates versus Numbers

Many of the best-known baseball statistics are counting statistics. These
include home runs, RBI, and strikeouts, where you simply count how many
times each thing happened. By contrast, batting average and on-base per-
centage are rate statistics computed as ratios. Batting average, for example,
is measured in hits per at bat. In general, counting statistics measure total
production and are highly influenced by how often a player is in position to
achieve the statistic. For example, a basketball player could be a great shooter
but will not score much if most of the shots are taken by teammates. Rate
statistics are more about efficiency, typically measuring how often an event
happens compared to the number of opportunities.
The most popular basketball statistic is points scored, which is a counting
statistic. While points per game is a rate statistic, the number of games played
is not a good representation of the player’s opportunities. Points per minute
is fairer to the player who gets limited minutes per game. Computing points
per 48 minutes puts the statistic on a familiar scale where 30 is a large value.
128 Sports Math

Example 7.3 Player A scored 28 points in 40 minutes and player B scored

12 points in 14 minutes. Compute points per 48 minutes for each player.
28 pts
Solution. Player A scored 40 min = 0.7 points per minute. Multiply by 48
12
minutes to get 33.6 points per 48 minutes. Player B scored 14 (48) = 41.1
points per 48 minutes. If player B had maintained the same scoring rate for
48 minutes, he or she would have scored 41 points.

Player B scored points at a faster rate than player A, given the time on the
court. Example 7.3 shows some of the advantages and disadvantages of rate
statistics. Knowing that a player scored 12 points might disguise the impact
that player B had on the game. However, you might not want to use points
per 48 minutes to conclude that B had a better game than A. In general,
players with limited time can post higher rates than players who would have
to maintain that high rate for a long period of time.
Suppose that two basketball teams each grab 10 offensive rebounds. Based
on this counting statistic, you would not see an advantage for either side.
Example 7.4 shows where a rate statistic uncovers a strength that could have
been missed.

Example 7.4 Team A made 45 out of 80 shots, while team B made 35 out
of 80 shots. Both teams have 10 offensive rebounds. Compute the offensive
rebounding rate for each team.
10
Solution. Team A missed 35 shots, and got offensive rebounds on 35 = 0.27
rebounds per opportunity. Team B missed 45 shots, and got offensive rebounds
on 10
45 = 0.22 rebounds per opportunity. Team A did a better job of rebounding
its own misses.

Persistence and Reliability

Many statistics are designed to measure a particular skill. In basketball,
shooting is measured by shooting percentage, rebounding by total rebounds
or rebounding percentage (as in Example 7.4), passing by assists, and so on.
Is there any way to tell whether a statistic actually captures a player’s ability?
One way of analyzing what a statistic measures is to check for consistency,
what is variously called persistence or reliability.
Suppose that we invent a new statistic that is designed to measure clutch
play, the ability to perform well at the most important times. We compile
our list of top 10 clutch players each year, but nobody ever makes the list
more than once. Year after year, players’ clutch ratings change in a seemingly
random pattern. There are two possibilities. One is that there is no such
thing as clutch play; there may not be a skill there to measure. The other
 Saber- and Other Metrics 129

possibility is that our statistic does not measure what we want. Our statistic
may be influenced by other skills at cross-purposes to clutch play, or our
statistic might be allowing in too much noise. Either way, we need to modify
or abandon our statistic.
A good skill statistic will be stable, in that a player or team who rates
highly one year should usually rate highly the next year. There are injuries
and other influences that will cause some change, but generally the best ratings
should persist. A way of testing this is to compile the statistics for consecutive
years for all players or teams who played both years. A high correlation be-
tween consecutive years is a good sign. Correlation is a measure of the linear
relationship between two variables. If there is an exact linear relationship (for
example, next year’s value is exactly the same as this year’s, or exactly half),
then the correlation equals 1, and we have high confidence that we can use
this year’s value to predict what next year’s value will be. If the two variables
are not related at all, the correlation will equal 0, and we have zero predictive
ability. (The converse is not true, so a correlation near 0 does not prove that
the variables are unrelated.) If the variables are related through a line with
negative slope (so large values one year predict small values the next year),
the correlation is −1. The correlation is always between −1 and 1.
A technical definition of correlation may aid those familiar with vector
analysis. If the n values of variable x are collected in a vector a and the n
values of variable y are collected in a vector b, then the correlation between
x and y equals the dot product of a and b divided by the magnitudes of a
and b. Stated differently, the correlation is the cosine of the angle between the
vectors: if the vectors line up in the same direction, the correlation is 1. The
opposite direction gives a correlation of −1. A correlation of 0 follows from a
90-degree angle, so that the vectors are perpendicular (or orthogonal, as we
like to say).
Baseball Between the Numbers, from the Baseball Prospectus group, com-
piled such year-to-year correlations for a variety of pitching statistics. Strike-
outs per batter faced (note the use of a rate statistic) had a year-to-year
correlation (using data from 1972 to 2004) of 0.790. Walks per batter was
less consistent at 0.676, then hits allowed per batter at 0.499, home runs per
batter at 0.470, ERA (earned run average) at 0.380, and winning percentage
at 0.204. The takeaway is not that winning percentage for pitchers is useless,
but that it clearly depends on many more factors than just the pitcher’s per-
formance. In fact, attempts to find pitching “winners” who win close games
by grit and iron will have failed as thoroughly as attempts to find clutch hit-
ters. Strikeouts, on the other hand, are very much under the control of the
pitcher and the high correlation shows that high (or low) strikeout rates tend
to be maintained by pitchers year after year. It makes sense to call a pitcher a
“strikeout pitcher” or a “groundball pitcher” (percentage of ground balls has
a year-to-year correlation of 0.807), but calling a pitcher a winner is not very
meaningful.
130 Sports Math

On the Defensive
There is an old sports adage that “defense wins championships.” We will
see in the exercises that this is not especially true. It may be that the adage
is mostly a nod to players who play good defense in sports where defensive
ability is hard to measure. Baseball is the major sport in which defense is close
to an individual endeavor, so we should be able to evaluate defensive ability in
baseball. And yet, we have the case of Derek Jeter, who won five Gold Glove
Awards as the best shortstop in the American League, but is consistently
rated as below average by sophisticated defensive ratings.
Jeter typically had a good fielding percentage. This is the original fielding
statistic, and is defined as the ratio of successful plays (assists plus putouts) to
chances (assists plus putouts plus errors). Errors are somewhat of a judgment
call, but require the fielder to get his glove on the ball. A slow fielder who
can’t even reach a ground or fly ball is not penalized, while a speedy fielder
who tracks down a ball that nobody else would get to is penalized if he then
makes an error on, for example, the throw. The effect can be seen in Example
7.5.

Example 7.5 Player A is a shortstop with 273 putouts, 392 assists, and 13
errors. Compute his fielding percentage. Player B is a little faster and reaches
28 more balls, leading to 27 more assists and 1 more error. Compute his
fielding percentage. Determine how many assists are needed to compensate
for one error.
273+392
Solution. Player A has a fielding percentage of 273+392+13 = 665
678 = 0.9808.
Adding 27 more plays to the numerator and 28 more chances to the denomi-
nator, his fielding percentage is now 692
706 = 0.9802, which is less than before.
716
Adding another 24 successful plays gives a fielding percentage of 730 = 0.9808.
That is, player A would need to make 51 successful plays to offset one error
in the fielding percentage statistic.

The question to ask is whether you would rather have player B or A. Most
would take B, even though his fielding percentage is worse. Player B made 27
more successful plays, a full game worth of outs. Sometimes runners advance
more on errors than on base hits, but often the impact of an error is identical
to that of a hit. Would you trade 27 outs for one hit? The starting numbers
are, perhaps, extreme but are taken from Derek Jeter’s 2004 season.
An early attempt at improving on fielding percentage is Bill James’s
Range Factor, which equals putouts plus assists divided by games played.
The simple idea is that the more plays you make, the better you are. An error
is no different than being too slow to get to the ball in the first place. To take
one example, in 2004 Derek Jeter won a Gold Glove at shortstop. His fielding
percentage of 0.9808 was fourth in the league among shortstops who played
more than 100 games, and his 273 putouts led the league. Yet, his Range
 Saber- and Other Metrics 131

Factor of 4.46 plays per 9 innings was below the league average of 4.56 plays
per 9 innings.
With new tracking data, discussed in Chapter 10, major improvements are
being made in defensive statistics.

Plus and Minus

In many sports, defense is very hard to measure because individual re-
sponsibilities are impossible to define. A basketball player gets a layup off of
a pick-and-roll. Does the fault lie with the player “guarding” this player? Per-
haps his or her role was to stop the dribbler and the pass was open because
the other defender did not switch to the roll player fast enough. But, where
was the help defense? And which side was the help to come from? It may be
impossible to assign blame correctly.
These types of plays in basketball and most of the action in hockey, foot-
ball, and soccer are so fluid that they are hard to evaluate. An idea that has
gained some traction is the plus-minus statistic, which attempts to divide
credit or blame for points. The basic form of the plus-minus for a player is
the net score (team points for minus points against) while that player is in
action (on the court, on the ice, and so on). If a five-person group outscores
the opposition by two, then each of the five people scores +2, while each of
the opponents (assuming no substitutions) scores −2.
The obvious flaw to the plus-minus is that a player on a bad team will
almost surely have a negative rating, no matter how good the player is. Brian
Burke famously expressed this in more colorful language at the 2014 Sloan
Sports Analytics Conference. (The session had the impish title “Hockey An-
alytics: Out of the Ice Age.”) A partial fix is to compute the difference of the
team’s net score when the player is in the game and out of the game. A player
on a bad team gets a +5 rating if the team is “only” outscored by 8 points
when the player is in the game compared to being outscored by 13 when the
player is out of the game. As a method for evaluating individual players, this
can still be unsatisfying. If a mediocre player is always put in the game by the
coach when the team superstar is playing, the mediocre player will look good.
If the lineups vary enough, there are ratings methods that can assign indi-
vidual contributions of players to each lineup. For example, in basketball each
five-person group can be treated as a distinct team, and a rating method such
as those discussed in Chapter 5 can be used to rate each team. An individual
player’s rating could be the sum of the group ratings for all of the groups to
which that player belongs, weighted by the number of minutes that the groups
played together and compared to the overall team rating.
Plus-minus ratings are used by coaches to evaluate the effectiveness of
different groupings of players.
132 Sports Math

Park Factors
A basketball court is 94 feet long, a hockey rink is 200 feet long, a tennis
court is 78 feet long, and so on. By contrast, baseball and golf playing fields
have dramatically different sizes and shapes. Further, outdoor sports can be
strongly influenced by environmental issues such as temperature and altitude.
For this reason, baseball in particular has a need for corrections for the site
of events. The statistics for someone playing half of his games in Coors Field
can and do look different from someone whose home field is Safeco Park in
Seattle.
The proper way to correct for home field in baseball is not entirely simple.
Suppose that 50% more home runs are hit in Fenway Park than in any other
stadium. Before you conclude that Fenway Park is 50% easier to hit home
runs in than average, think about which teams play there. If the Red Sox are
loaded with home run hitters, then the extra home runs in Fenway could be a
result of Red Sox ability instead of the park. Also, the result could be due to
the generosity of the Red Sox pitchers. We need to factor in Red Sox games
both in Fenway and away from Fenway to get a read on how much the park
affects home runs. The following gives a simple method of computing park
factors.
If Hh is the number of home runs that team A hit at home, Ha is the
number of home runs that team A allowed at home, Rh is the number of
home runs that team A hit on the road, and Ra is the number of home runs
that team A allowed on the road, then team A’s park factor for home runs is
given by
(Hh + Ha )/Hg
PF =
(Rh + Ra )/Rg
where Hg is the number of home games that team A played and Rg is the
number of road games played. In a full season, Hg = Rg = 81, and the park
factor simplifies to the number of home runs hit (by either team) in team
A’s home games divided by the number of home runs hit in team A’s road
games. In 2014, Yankee Stadium had the largest park factor for home runs
at 1.47 and AT&T Park in San Francisco had the lowest at 0.68. (Data from
espn.com.)
You may be surprised that Coors Field did not have the largest home run
park factor (it was second at 1.39). The reason indicates why park factors for
other statistics need to be computed. With outfield fence dimensions of 347
feet to left, 410 feet to center, and 350 feet to right, Coors Field has one of
the largest fields in baseball. The fences need to be pushed back to reduce
the number of home runs. However, the spacious outfield allows more balls to
fall for hits: the 2014 Coors Field park factor for hits is 1.32 (second place is
1.08). Another factor is the amount of foul ground: parks with the seats on
top of the field do not allow fielders to catch many pop fouls for outs, giving
 Saber- and Other Metrics 133

hitters extra pitches to face. When you put it all together, Coors Field has the
largest runs park factor at 1.50 (50% more runs are scored at Coors Field!),
with Chase Field in Phoenix second at 1.15 and Safeco Park lowest at 0.82.
With this method of computing park factors, a player’s statistics can be
park-adjusted in a number of ways. Since Coors Field multiplies home run
totals by 1.39, you might think that a Colorado player should have his home
run total divided by 1.39. This, however, assumes that all of the player’s games
are at Coors. Instead, suppose a player played half of his games in Coors Field
and the other half in parks for which the average park factor is 1. Then half
of his home runs should be divided by 1.39. In general, for a player whose
2x
home field has park factor pf , the park-corrected value should be 1+pf where
x is the uncorrected statistic. In the case of Coors Field home runs, notice
2 1
that 1+pf = 1.195 . Instead of dividing out the 39% increase that Coors Field
provides, we divide out the 19.5% increase that playing half of his games in
Coors Field provides.

Example 7.6 In 2014, Nelson Cruz led the American League with 40 home
runs for Baltimore. Chris Carter of Houston was second with 37. Given home
run park factors of 0.936 for Baltimore and 1.173 for Houston, compute the
park-adjusted home run numbers. In 2014, Clayton Kershaw led the National
League in ERA at 1.77 for Los Angeles. Johnny Cueto of Cincinnati was
second at 2.25. Given run park factors of 0.907 for Los Angeles and 0.963 for
Cincinnati, compute the park-adjusted ERA numbers.
Solution. We can see that Cruz played in a tough park for home runs, while
80
Carter had it relatively easy. The park-adjusted value for Cruz is 1.936 =
74
41, while the park-adjusted value for Carter is 2.173 = 34. Cruz earned his
home run championship. For the pitchers, both Kershaw and Cueto played in
3.54
pitcher-friendly parks. The park-adjusted value for Kershaw is 1.907 = 1.86
4.5
and the park-adjusted value for Cueto is 1.963 = 2.29. Kershaw still has a
sizable lead, but the lead is not as large.

You should object to our use of the same adjustment model for pitchers and
batters. The assumption behind the model is that half the games are played
at home and half on the road at parks that have an average park factor of
1. For batters who play every day, this is not a bad assumption, unless their
team plays in a division with unusually high or low park factors. Pitchers, who
only pitch in every fifth game or so, could easily have an unequal distribution
of parks pitched in. It would be more accurate in all cases to weight the
adjustments by how many games are played in each park.
You might consider a further tweak to the park factor model, trying to
account for different parks being better or worse for different players. Fenway
Park is great for lefthanded line drive hitters who have would-be outs bounce
off the Green Monster for doubles, while righthanded line drive hitters whose
would-be home runs bounce off the Monster are hurt by Fenway Park. (The
134 Sports Math

doubles park factor for Fenway Park in 2014 was 1.523, by far the largest in
the major leagues.)

Four Factors, Fenwick, and Football

In the exercises, you will be asked to explore questions about a variety of
sports, with suggested websites to find relevant information. Since the text to
this point has been baseball-heavy, in this section we look at a couple of ideas
from other sports.
Dean Oliver’s Four Factors evaluate basketball team performances. For
our purposes, they also illustrate an important guide to developing ratings.
Among statistics that are readily available, which ones identify the better
team? Points scored is the obvious answer, but it’s not a helpful answer.
Another unhelpful, though correct, answer is “all of them.” To give an example
of why this is unhelpful, suppose you have numbers for offensive rebounds,
defensive rebounds, and total rebounds. You don’t want to use all three in
your analysis: if you have two of them, it is not hard to figure out the third.
The more interesting question is whether offensive rebounds are more or less
important than defensive rebounds. And, by the way, we should remember to
frame the statistics as rebound rates and not just raw numbers.
In order of importance, Oliver’s four factors are shooting, turnovers, re-
bounding, and free throws. There is nothing surprising here, but the trick
is to define each category in a way that extracts useful information without
duplicating information from other categories. As discussed in Example 7.4,
shooting and rebounding can overlap. Here is one version of the Four Factors.
Shooting is measured by effective field goal percentage where 3-point
shots made are given 50% more credit (since they are worth 50% more points).
Then we compute S = FGM+0.5TPF FGA where the team attempted FGA field
goals and made FGM field goals, of which TPF were three-pointers. A team’s
shooting rating for a game is the difference in S-values for that team and its
opponent.
Turnovers is a rate statistic, with the denominator being the number of
possessions. Since the number of possessions is not a statistic that is normally
available, we use the estimate FGA − OREB + 0.4FTA + TO where OREB is
offensive rebounds and free throws attempted (FTA) is multiplied by 0.4 to
compensate for multiple free throws per possession. Typically, 2 free throws
constitute a possession, but “and-one” free throws do not represent a new
possession and sometimes a player shoots three free throws. On average, 0.4
(some use 0.44) seems to work well. The team’s turnover rating is the difference
TO
in its rating and its opponent’s rating of T = FGA−OREB+0.4FTA+TO .
Rebounding is a rate statistic, with offensive rebounding measured by
OREB
OREB+DREBopp where DREBopp is the number of defensive rebounds by the
 Saber- and Other Metrics 135

opponents. The denominator could be slightly different from the number of

missed shots FGA−FGM. Defensive rebounding is the complement of the op-
ponents’ offensive rebounding rating, so computing the difference in offensive
rebound ratings accounts for all rebounds.
Finally, free throws are measured as a rate statistic. Instead of looking at
free throws per possession, Oliver divides free throws made by the number of
field goal attempts to compute FTM
FGA .

Example 7.7 In a 2015 playoff game, San Antonio made 42 of 91 field goal
attempts, 8 of them three-pointers, made 19 free throws, had 10 offensive and
38 defensive rebounds, and had 9 turnovers. Los Angeles made 39 of 92 field
goal attempts, 9 of them three-pointers, made 20 free throws, had 16 offensive
and 39 defensive rebounds, and had 11 turnovers. Compute the Four Factors
for each team.
Solution (1) San Antonio’s effective field goal percentage is 42+4
91 = .505, while
Los Angeles’ is 39+4.5
92 = .473. San Antonio’s shooting rating is .505 − .473 =
.032, so San Antonio shot better. (Los Angeles’ rating will be the negative
of San Antonio’s; in this case, −.032.) (2) San Antonio’s turnover rate is
9 11
91−10+10.4+9 = .0896, while Los Angeles’ rate is 92−16+14.8+11 = .1081. San
Antonio’s turnover rating is .0896 − .1081 = −.0185, so again San Antonio’s
rating is better. By the way, notice that the denominators are similar: 100.4
for San Antonio and 101.8 for Los Angeles. Since the denominator represents
possessions, the numbers should be within one of each other. (3) San Antonio’s
10
offensive rebounding rate is 49 = .204 and Los Angeles’ is 16 54 = .296. San
Antonio’s rebounding rating is .204 − .296 = −.092, so San Antonio lost the
rebounding battle. (4) San Antonio’s free throw rate is 19 91 = .209 to Los
Angeles’ 20
92 = .217. San Antonio’s free throw rating is .209 − .217 = −.008, so
San Antonio narrowly lost the free throw battle.

Each team won two of the four ratings, indicating that the game was close.
In fact, San Antonio won 111-107.
Wayne Winston’s book Mathletics notes that correlations between any
two of the factors are close to zero. Mathematically, we would say that the
Four Factors are “orthogonal.” This means that the information contained in
one factor does not duplicate the information in any other factor. This is a
desirable, and somewhat rare, property.
Hockey analytics are not as advanced as baseball or basketball analytics.
Hockey is very fluid with numerous line changes, making it difficult to isolate
individual performances. With few goals scored, statisticians have little scoring
data to work with. Hockey was an early adopter of the plus-minus statistic.
The flaws of plus-minus are magnified in hockey, since goals are infrequent
and are often scored on power plays. Penalty killers will have bad plus-minus
numbers, unless adjustments are made for man-up situations.
An important finding in hockey is that shots taken have more predictive
power than goals. That is, when predicting future success you are more likely
136 Sports Math

to be right if you use the number of shots taken in past games than if you
use the number of goals scored in past games. Good teams create more shots
than their opponents. Bad luck (shots that barely miss or goalies that make
amazing saves) can keep a good team from scoring goals, but the number
of shots is a good measure of ability. This leads to two statistics, which are
named for the people who popularized them.
Fenwick measures the number of shots. As just noted, this is a good predic-
tor of future success. To be exact, start by counting shots (on goal or not). A
team’s Fenwick rating is given by shots for minus shots against. A player’s
Fenwick rating is shots for minus shots against while that player is on the
ice. Basically, the old plus-minus statistic based on goals has been updated
to the better statistic of shots. To make it a rate statistic, compute the Fen-
wick percentage equal to shots for divided by total shots. The Corsi rating
is identical, except that blocked shots (where a defenseman blocks the puck
before it reaches the net) are included in the counts.

Example 7.8 While Patrick Kane is on the ice, his Chicago team takes 12
shots, of which 3 are blocked by defensemen. The opposition takes 6 shots, 1
of which is blocked. Compute Kane’s Fenwick and Corsi percentages.
Solution For Fenwick, 9 of the Chicago shots and 5 of the opponents’ shots
count. Kane’s Fenwick rating is 9 − 5 = +4 and his Fenwick percentage is
9
9+5 = .643. For the percentage, above .5 is better than average. For the Corsi
12
rating, count all shots. Kane’s Corsi percentage is 12+6 = .667.

Football (the American version) has a long history of statistics. In 2014,

DeMarco Murray led the NFL in rushing with 1845 yards. Drew Brees led the
NFL in passing with 4952 yards, while Andrew Luck led the league with 40
touchdown passes. These are all counting statistics. That does not make them
worthless, but it probably does indicate the difficulties of isolating individual
contributions in football. For example, suppose you learn that Brees just threw
a 75-yard touchdown pass. You might envision a perfectly thrown long pass,
but it may have been a 5-yard slant pattern that the receiver broke for a
touchdown. How much credit does Brees deserve?
An early attempt at accounting for all of the credits and debits in a quar-
terback’s performance is the NFL’s Quarterback Rating Formula. An abbre-
viation of PR (for “Passer Rating”) will be used. Like basketball’s Four Fac-
tors, there are four components of the formula: completion percentage, average
yards gained per attempt, touchdowns per attempt, and interceptions per at-
tempt. Like the NBA’s Four Factors, all components are rates. According to
nfl.com, the following four calculations start the process for a quarterback with
c completions in a attempts for y yards with t touchdowns and n intercep-
tions. Compute (1) (100c/a − 30) ∗ 0.05, (2) (y/a − 3) ∗ 0.25, (3) (100t/a) ∗ 0.2,
(4) 2.375 − (100n/a) ∗ 0.25. For each of the four components, replace numbers
greater than 2.375 with 2.375 and numbers less than 0 with 0. Add the four
 Saber- and Other Metrics 137

numbers (each of which is between 0 and 2.375), multiply by 100/6, and you
have PR.
You may be glad to know that we can do some work to make this clearer.

Example 7.9 In 2014, Drew Brees completed 456 of 659 passes for 4952
yards, with 33 touchdowns and 17 interceptions. Compute his Quarterback
Rating.
Solution The four calculations are (1) (100 ∗ 456/659 − 30) ∗ 0.05 = 1.960,
(2) (4952/659 − 3) ∗ 0.25 = 1.129, (3) (100 ∗ 33/659) ∗ 0.2 = 1.002 and (4)
2.375 − (100 ∗ 17/659) ∗ 0.25 = 1.730. Since all four numbers are between 0
and 2.375, no adjustments are necessary. Combine the numbers to get PR =
(1.960 + 1.129 + 1.002 + 1.73) ∗ 100/6 = 97.02.

This rating placed Brees 6th among quarterbacks who threw at least 100
passes in 2014. (Tony Romo was first with a rating of 113.2.) The rating is not
easy to compute, and harder to logically decipher. The four categories make
sense, but where do those constants come from? Let’s do some algebra and
see what pops out. Start with

(100c/a − 30) ∗ 0.05 + (y/a − 3) ∗ 0.25 + (100t/a) ∗ 0.2 + 2.375 − (100n/a) ∗ 0.25

and multiply to get 5c/a − 1.5 + 0.25y/a − 0.75 + 20t/a + 2.375 − 25n/a.
Combine like terms to get 0.125 + 5c+0.25y+20t−25n
a . The Quarterback Rating
can be rewritten as
5c + 0.25y + 20t − 25n 100
PR = (0.125 + )∗
a 6
for cases in which each of the four components is between 0 and 2.375.
We can now see how the different variables are valued. One touchdown
is equivalent to 20/0.25 = 80 yards, one interception is equivalent to 100
yards, and one completion is equivalent to 20 yards. Where did these numbers
come from? We will find better constants later in this chapter, and a better
quarterback rating system in Chapter 10.

Evaluation and Prediction

The word “prediction” is used in different ways in this chapter. The num-
bers of runs a baseball team scored and allowed should predict the team’s
winning percentage. This means that a team that scored 665 runs and allowed
614 in 162 games should have won about 87.5 games. The games have already
been played, so this is not a prediction about what will happen in the future.
By contrast, a team that after 40 games has scored 181 runs and allowed 146
138 Sports Math

runs can be predicted to win 98.1 games out of 162. This is predicting what
will happen in the future.
Sports analytics is interested in both types of predictions. The first type,
using what happened with one statistic to estimate what happened with an-
other statistic, is used to validate statistics to show that they have a strong
relationship to what they are intended to measure. The second type is of more
interest, which is to use statistics derived from past performance to predict
performance in the future. Since we do not have data about the future, this
second type of prediction is more difficult.
There are different ways to simulate future predictions. One way is to use
data from two years ago to try to predict the results from one year ago. The
resulting model could then be used to predict how many goals a given team
will score next year, using this year’s goals or this year’s shots. A slightly
different question is of a more theoretical nature and was alluded to earlier.
Which statistics are most indicative of high quality play? The measure of
“indicative” is an ability to predict results. We split our data in half, and
use the first half of the data to create a good prediction of the second half
of the data. Dividing the data chronologically can be problematic, if trades
and injuries change a team dramatically in the second half of the season. You
can, instead, divide the data in half by putting the odd-numbered games in
one half and the even-numbered games in the other half. You still have half of
the season to “predict” but it is less likely that the data will be contaminated
by large changes in the team. It is in this sense that past numbers of shots in
hockey predict future goals better than past goals do.

Regression to the Mean

If a player has a break-out season, performing far better than in the past,
what should you predict for the future? Among the options are to project an
even higher level, the same level, or a lower level of performance. A number
of variables could go into this projection (age being an important one), but in
general the safest bet is to predict the player to drop back closer to an average
performance. This is called regression to the mean.
Using team passing yards from nfl.com, a linear equation for using yards
p13 in 2013 to predict yards p14 in 2014 is

p14 ≈ p13 − 0.33(p13 − 3770).

This equation is the “best” predictor in the sense that out of all equations of
the form a p13 + b this equation’s predictions are closest to the actual 2014
values (in the “least squares” sense discussed in the next section). The signif-
icance of the constant 3770 is that it is the average passing yards for a team
 Saber- and Other Metrics 139

in 2013. So, our prediction is to take the previous year’s total and remove
one-third of the difference between the previous year’s total and the average
total. The prediction for next year is closer to the mean than last year’s value
was.

Example 7.10 In 2013, Denver passed for 5444 yards and Buffalo for 3103
yards. Use the regression equation to predict passing totals for Denver and
Buffalo for 2014.
Solution For Denver, the equation gives us 5444 − .33(5444 − 3770) = 4892
yards, a considerable reduction. In fact, Denver had 4661 passing yards in
2014. For Buffalo, the equation gives us 3103 − .33(3103 − 3770) = 3323 yards,
a nice increase from the below-average total in 2013. In fact, Buffalo had 3614
yards.

Regression to the mean explains a number of sports superstitions. A rookie

who has a great season is subject to a “sophomore slump” that lowers pro-
duction in season two. Unless the player is truly outstanding, regression to
the mean explains a second season that is closer to average than the first sea-
son. Players and teams appear on the cover of Sports Illustrated shortly after
outstanding performances. An immediate drop in production may be due to
regression to the mean instead of the “SI cover jinx.” A team fires its manager
or coach and immediately starts performing better. Regression to the mean
predicts that a bad team will improve, closer to average, whether the manager
is fired or not.

Linear Weights: A Prelude to WAR

We evaluated batting average and on-base percentage using the line that
best fits the data in Figure 7.3. In Example 7.10, the equation of the best
fit line is used to predict passing yards for NFL teams. Both are examples of
linear regression, which we develop more fully in this section.
Figure 7.6 shows the average score on the
PGA Tour in 2013 as a function of hole length
for three distances. The score increases as the
distance increases, in what appears to be a lin-
ear fashion. If we had an equation for the best
fit line, we could predict average scores for any
distance. By “best fit” we mean, as always, that
the sum of the squares of errors is minimized. FIGURE 7.6: Score as a
Suppose the line has equation y = mx + b for Function of Distance
constants m and x. For a hole of length x = 420
yards, the predicted score is m ∗ 420 + b and the actual score is 4.01. The
140 Sports Math

square of the error is (420m + b − 4.01)2 . The calculus box below shows how
to find m and b to minimize SSE. We get a = 0.00185 and b = 3.24.
The NFL Quarterback Rating combines four statistics into a single rating.
The weights (the constants multiplying the variables) are mysterious, however.
Ideally, the weights would be chosen to minimize the sum of the squares of
errors for points scored. This is what multiple regression does. In this case,
we want the linear combination of the four variables c/a, y/a, t/a, and n/a
that minimizes SSE for team points scored. Each team has c completions in
a attempts for y yards with t touchdowns and n interceptions. Using team
totals for 2013 and 2014 from nfl.com, we get the following regression formula
(scaled so that the coefficient of y/a is the same 0.25 as in the NFL system):

0.51 + 0.32(c/a) + 0.25(y/a) + 25.29(t/a) − 11.47(n/a).

The constants here are not totally different from those used by the NFL.
The largest difference is that the NFL weights completions far too heavily (5,
versus the optimal 0.32). The NFL system was created in 1973 when most
passes were thrown far down field, making completions more important. The
modern passing game of short passes makes completions less significant, and
this shows up in the reduced weight for completions. Interestingly, touchdowns
are weighted almost exactly the same and interceptions are weighted only
half what they were in 1973. Despite its obscure presentation, PR (the NFL
Quarterback Rating) is at heart a regression of four basic statistics against
points scored.
Multiple regression is also used to create a prediction of runs scored in
baseball from basic counting statistics like singles (1B), doubles (2B), and
so on. The model is often called Linear Weights. Using values from Tom
Tango’s website tangotiger.net, one version of Linear Weights is given by

1.409(HR) + 1.063(3B) + 0.764(2B) + 0.474(1B) + 0.330(BB + HBP)+

0.195(SB) − 0.456(CS) − 0.299(outs).

The coefficients in the Linear Weights formula give us information about

the relative worth of events. We see that a walk is almost as productive as
a single, giving some validation to the use of on-base percentage instead of
batting average. Linear Weights is not kind to base stealers. Practitioners
of “Moneyball” de-emphasize base stealing, based on calculations like the
following.

Example 7.11 A player steals 14 bases in 21 attempts. Based on Linear

Weight, how many runs has he created? What is the success rate needed for
a base stealer to break even?
Solution The player has SB = 14 and CS = 7. The runs created equal 0.195 ∗
14 − 0.456 ∗ 7 = −0.46. The player actually cost his team about half a run.
To break even, we need 0.195(SB)=0.456(CS) or CS = 195 456 SB. This leads to
SB SB
a success rate of SB+CS = SB+0.428SB = 0.700.
 Saber- and Other Metrics 141

By these numbers, a base stealer must be successful more than 70% of the
time, or he is hurting his team! In 2014, major league totals were 2764 steals
in 3799 attempts, a 72.7% success rate. On the whole, teams have learned to
be smart about how often to attempt to steal bases.
Linear Weights can be used to rate players’ offensive contributions. Sub-
stitute in a player’s yearly totals and you get an estimate of the number of
runs the player created. Since we can translate runs into wins (approximately
10 runs for one win), this allows us to estimate how many wins a player
creates. This is the concept behind WAR (Wins Above Replacement), an all-
encompassing statistic that attempts to measure the full range of a player’s
contributions (batting, fielding, baserunning, pitching) in terms of wins cre-
ated compared to an average replacement player at that position.

Calculus Box: Linear Regression

The details of finding a (best fit) regression line are given here. The data
points in Figure 7.6 are (390, 3.96), (420, 4.01), and (450, 4.071). The errors
for predictions from the line y = mx + b are 390m + b − 3.96, 420m + b − 4.01,
and 450m + b − 4.071. Then

SSE = (390m + b − 3.96)2 + (420m + b − 4.01)2 + (450m + b − 4.071)2

which we want to minimize. In calculus, at its minimum point a function

has all first order partial derivatives either equal to zero or nonexistent. The
partial derivative with respect to m gives the equation

2(390m+b−3.96)∗390+2(420m+b−4.01)∗420+2(450m+b−4.071)∗450 = 0

and the partial derivative with respect to b gives the equation

2(390m + b − 3.96) + 2(420m + b − 4.01) + 2(450m + b − 4.071) = 0.

We solve the two equations and two unknowns through substitution, elimi-
nation, or matrix inversion. The equations are 531000m+1260b = 5060.55 and
1260m+3b = 12.041. Solving the second equation for b gives b = 4.014−420m.
Substituting into the first equations gives 531000m + 1260(4.014 − 420m) =
5060.55 or 1800m = 3.33 and then m = 0.00185. Then b = 4.014 − 420 ∗
.00185 = 3.24.

Roger Maris and the Hall of Fame

I grew up as a Roger Maris fan. (Mickey Mantle lived nearby and I went
to high school with his sons, so I am also a fan of The Mick.) There is not a
142 Sports Math

large movement to get Roger Maris into the Hall of Fame, beyond the extent
to which his 61 home runs in 1961 are already memorialized. However, Tom
Clavin and Danny Peary’s biography of Maris makes an interesting case for
his abilities as a baserunner and fielder in addition to his home run hitting.
Here are some basic Maris stats: lifetime batting average of .260 (not good),
275 home runs (good, not great, given an injury-shortened 12-year career),
two MVP awards (excellent). The detailed statistics that will be discussed in
Chapter 10 are not available for players from the 1960s, but play-by-play data
is fairly complete going back to 1940. Using data from baseball-reference.com,
a statistic called base-out runs added estimates the number of runs that a
batter or baserunner adds compared to average. Maris ranks 151st for career
totals, surrounded by Steve Garvey, Bernie Williams, Hall of Famer Enos
Slaughter, and Ken Griffey. As a fielder, Maris was a top-five right-fielder in
Range Factor in each of the years that he played right field primarily (he
often filled in as a center fielder). An estimate of the total runs saved as a
right-fielder has Maris ranked 29th among right-fielders.
All in all, in career WAR value Roger Maris ranks 351st, between Kirk
Gibson and Chili Davis. It could be that the version of WAR used by baseball-
reference.com underestimates Maris’s contributions to his team, but the evi-
dence here is for a career that was outstanding but, due to injury, falls short
of Hall of Fame status.

Now Trending
Not all useful information requires detailed statistical analysis. Often, you
can find interesting patterns with very simple tools. A time series is a se-
quence of values of some variable at discrete moments in time. Here, we look
at some year-by-year progressions to quantify basic changes in the way sports
are played.

FIGURE 7.7: Missing You More

 Saber- and Other Metrics 143

Figure 7.7 shows the alarming rate at which major league strikeouts have
increased since 2005. (Data from baseball-reference.com.) The increase is al-
most linear. However, it would be silly to use a best fit line to predict future
strikeouts. Presumably, the players or rules-makers will decide that this trend
is harmful to the game and do something about it. The figure is a good test
case for interpreting a time series. The ups and downs from 2000-2005 are
likely a product of “random” variation, while the lengthy and steady rise
from 2005 to 2014 represents a real trend that deserves attention.
Figure 7.8 shows the average passing yards per team in the NFL. (Data
from pro-football-reference.com.) Passing yards has also shown a steady in-
crease since 2005.

FIGURE 7.8: A Passing Fancy

Figure 7.9 shows the average number of three-point shots attempted per
game. (Data from basketball-reference.com.) You can clearly see the increased
usage of the shot over time since its introduction in 1979 (when a mere 2.8
shots were attempted per game). Note that the legend “2014” refers to the
2014-15 season. Analysts claim that layups and three-point shots are the most
efficient shots in the game, and teams are clearly using the three-point shot
more often.

FIGURE 7.9: The Gang is Shooting Threes

Figure 7.10 shows that a trend that is “well known” is not occurring.
There has been widespread discussion that scoring in soccer is diminishing,
144 Sports Math

with doomsday predictions of 0-0 draws becoming the norm. Contrary to this
belief, the number of goals in the English Premier League has been (until the
2014/15 season) quite high. (Data from soccer-europe.com.)

FIGURE 7.10: Goals Scored

More trends will be explored in the exercises.

Exercises
In these exercises, T refers to thinking problems, conceptual problems requir-
ing no calculations. C refers to problems requiring significant calculations or
calculus. P refers to projects; these are ideas for further investigation (hints
and resources are at the book’s web site).

7.1 The 1983 Chicago Cubs scored 701 runs and allowed 719 runs. Use the Bill
James Pythagorean Method to predict the number of wins for the Cubs over a
162-game schedule.
7.2 The 1983 Cubs won 71 games. Explain why this is one reason for Bill James
to predict in his 1984 Baseball Abstract that the Cubs would do better in 1984.
(In fact, the Cubs won their division with 96 wins.)
7.3 The 2012 Baltimore Orioles won 93 games, scoring 712 runs and allowing
705. Compare actual wins to Pythagorean expected wins. Give reasons why
there might be such a large difference. Based on your explanation, would you
expect a similar difference in 2013?
7.4 The 2013 Orioles won 85 games, scoring 745 runs and allowing 709. Compare
actual wins to Pythagorean expected wins. How does this compare to the differ-
ence in 2012? Comment on the use of the term “luck” describing the differences.
7.5 Halfway through the 2014 season the standings in the National League West
were San Francisco in first with a winning proportion of .573, then Los Angeles
at .554, Colorado at .427, Arizona at .422, and San Diego at .420. Use the
runs scored/against figures for each team to compute the Pythagorean winning
proportions: San Francisco 338/304, Los Angeles 350/302, Colorado 405/427,
 Saber- and Other Metrics 145

Arizona 333/400, San Diego 240/300. Which team does the Pythagorean Method
predict to win the division? The season-ending winning proportions were .543,
.580, .407, .395, .475. Did the Pythagorean Method correctly predict the winner?
For these five teams, did the halfway proportions or the halfway Pythagorean
proportions better match the final winning proportions?
7.6 Look up the MLB standings for 2014 and identify one team with a large
positive “luck” value and one team with a large negative “luck” value. Track
that team’s record over ten years. Does that team have a predictable pattern of
positive or negative luck?
7.7 On June 11, 2015, the Cleveland Cavaliers narrowly out-rebounded the
Golden State Warriors 49-44. Cleveland made 29 out of 88 shots and had 16
offensive rebounds, while Golden State made 36 of 77 shots and had 6 offen-
sive rebounds. Compute offensive rebounding rates and describe the rebounding
difference.
7.8 The average number of runs scored by major league teams in 2000 was 832.
(baseball-reference.com) In succeeding years, the average was 773, 747, 766, 779,
744, 787, 777, 753, 747, 710, 694, 701, 675, and 659 (in 2014). Plot the time series
and discuss if there is a long-term trend for run production to decrease.
7.9 Look up the average number of walks per team from 2000-2014, plot the time
series, and discuss any trends. Repeat for home runs and hits.
7.10 The average number of points scored by NFL teams in 2000 was 20.7. (pro-
football-reference.com) In succeeding years, the average was 20.2, 21.7, 20.8,
21.5, 20.6, 20.7, 21.7, 22.0, 21.5, 22.0, 22.2, 22.8, 23.4, and 22.6 (in 2014). Plot
the time series and discuss if there is a long-term trend for points to increase.
7.11 Look up the average number of rushing yards per team from 2000-2014, plot
the time series, and discuss any trends. Repeat for first downs.
7.12 The pace (average number of possessions per team) in the NBA teams in
2000-01 was 91.3. (basketball-reference.com) In succeeding years, the average
was 90.7, 91.0, 90.1, 90.9, 90.5, 91.9, 92.4, 91.7, 92.7, 92.1, 91.3, 92.0, 93.9, and
93.9 (in 2014-15). Plot the time series and discuss if there is a long-term trend
for the pace to increase.
7.13 Look up the average field goal percentage per team from 2000-2014, plot
the time series, and discuss any trends. Repeat for free throw percentage.
7.14 Look up the offensive and defensive ranks of teams in championship games,
use the ranks to label the team with the better offense and better defense, and
compare how often the better offense wins versus how often the better defense
wins. (a) NFL 1970-2014; (b) NBA 1970-2014; (c) NHL 1970-2014.
7.15 In 2014, Gold Glove-winning shortstop J.J. Hardy made 13 errors in 594
chances. Compute his fielding percentage. Find how many chances another fielder
could have handled to have the same fielding percentage with 12 errors. Which
shortstop would be more valuable?
7.16 In 1965, Hall of Fame shortstop Luis Aparicio made 20 errors in 697 chances.
Compute Aparicio’s fielding percentage and compare to J.J. Hardy’s in 2014.
Given that both played in 141 games (and both for Baltimore), which shortstop
do you think had the better year? What other information would be good to
have?
7.17 In 2014-15, the Golden State Warriors played the lineup Curry-Thompson-
Barnes-Green-Bogut for 812 minutes, outscoring the opponents by 358 points.
146 Sports Math

The lineup Curry-Thompson-Barnes-Green-Speights played for 201 minutes,

outscoring the opponents by 78 points. Compute the net points per minute for
each lineup, and discuss the relative worth of Bogut and Speights. What other
information would you want to know to make a better judgment?
7.18 In 2014, the home and away home run numbers for 81 games each are (a)
Colorado: Hh = 119, Rh = 67, Ha = 90, Ra = 83; (b) Boston: Hh = 49, Rh = 74,
Ha = 67, Ra = 87; (c) San Diego: Hh = 54, Rh = 55, Ha = 47, Ra = 70; (d)
Baltimore: Hh = 107, Rh = 104, Ha = 68, Ra = 83. Compute home run park
factors for each team.
7.19 In 2014, the home and away hits numbers for 81 games each are (a) Col-
orado: Hh = 924, Rh = 627, Ha = 825, Ra = 703; (b) Boston: Hh = 696,
Rh = 659, Ha = 760, Ra = 698; (c) San Diego: Hh = 597, Rh = 602, Ha = 598,
Ra = 702; (d) Baltimore: Hh = 705, Rh = 729, Ha = 670, Ra = 672. Compute
hits park factors for each team.
7.20 In 2013, Miguel Cabrera was chosen as MVP over Mike Trout. Their
stats are listed. Compute Linear Weights for each. Also, compute a similar
stat called weighted on-base percentage (wOBA) given by (.72 BB + .75
HBP + .90 1B + 1.24 2B + 1.56 3B + 1.95 HR)/PA. Discuss your results.
BB HBP 1B 2B 3B HR PA SB CS outs
Trout 110 9 115 39 9 27 716 33 7 399
Cabrera 90 5 122 26 1 44 652 3 0 362
7.21 Compute the Four Factors and discuss the fact that the Warriors won both
FGA FGM 3P FTM OREB DREB TO
games. (a) Warriors 88 39 10 20 11 37 12
Cavaliers 94 39 9 13 13 32 11
FGA FGM 3P FTM OREB DREB TO
(b) Warriors 85 37 13 18 7 32 9
Cavaliers 82 32 6 27 16 40 16
7.22 In 2014-15, Alex Ovechkin had 217 shots on goal, 109 missed shots, and
141 blocked shots in 1215 minutes. Compute Ovechkin’s Fenwick, Corsi, and
Fenwick per 60 minutes ratings. Repeat for Patrick Kane, who had 124 shots
on goal, 42 missed shots, and 54 blocked shots in 908 minutes. (Data from
hockeyanalysis.com.)
7.23 In 1966, Don Meredith completed 177 passes in 344 attempts for 2805 yards
with 24 touchdowns and 12 interceptions. Sonny Jurgensen completed 254 passes
in 436 attempts for 3209 yards with 28 touchdowns and 19 interceptions. (a)
Compute the NFL Quarterback Rating PR for each. (b) The NFL totals were
3149 completions in 6108 attempts for 37436 yards with 280 touchdowns and
318 interceptions. In 2014, the NFL totals were 11200 completions in 17879
attempts for 121247 yards with 807 touchdowns and 450 interceptions. Compute
the league PR ratings for the two years.
7.24 On 11/27/14, Tony Romo completed 18 passes in 29 attempts for 199 yards
with 0 touchdowns and 2 interceptions. On 12/21/14, he completed 18 passes in
20 attempts for 218 yards with 4 touchdowns and 0 interceptions. Compute his
PR for each game.
7.25 Determine the maximum PR for a game. Give an example of statistics that
achieve the maximum.
7.26 The best fit line for predicting 2014 rushing yards from 2013 rushing yards
 Saber- and Other Metrics 147

is approximately r14 = 1383 + .22r13 . Show that this gives (approximately) the
regression to the mean equation r14 = r13 − .78(r13 − 1780), where the average
rushing yards for a team in 2013 was 1780.
7.27 The best fit line for predicting 2014 rushing yards per attempt from 2013
rushing yards per attempt is approximately ra14 = 3.44 + .169ra13 . Show that
this gives (approximately) the regression to the mean equation ra14 = ra13 −
.831(ra13 − 4.2), where the average rushing yards per attempt for a team in 2013
was 4.2.
7.28 The regression to the mean equation has the form y = x − c(x − a) where
x is the previous value, a is the average value and c is a constant. Show that if
c = 1, the prediction y equals the average value; if c = 0, the prediction y is the
previous value. Based on this, does a persistent statistic correspond to a larger
or smaller value of c?
7.29 T Give several reasons why a team’s actual record could deviate from its
Pythagorean Method prediction. Discuss whether that trend should continue or
not; that is, are you describing a repeatable skill or luck?
7.30 T Figure 7.3 indicates that on-base percentage was more important for
teams in 2004 than batting average. Assuming that teams acted on this infor-
mation, what would be the impact in terms of playing time for players who do
or do not walk often? Explain why this might produce the neutral situation of
Figure 7.5.
7.31 T A small budget team like the Oakland A’s needs to find undervalued
players. Use Figures 7.3 and 7.5 to explain why the A’s could find high-OBP
players for cheap in 2004 but not in 2014. Comment on the “moving target”
nature of finding undervalued players.
7.32 T Describe a situation in which a counting statistic is a better representa-
tion of the value of a player than the corresponding rate statistic, and a situation
in which the rate statistic is better.
7.33 T Discuss the importance of a statistic being persistent.
7.34 T In golf, Strokes Gained putting is not especially persistent (correlation
of 0.44). Give at least two possible reasons having to do with the skill of putting.
7.35 T If a personal statistic is not persistent for players who change teams or
coaches, discuss whether this statistic measures a skill or not.
7.36 T Give at least two possible reasons why winning percentage for a pitcher
is less persistent than strikeouts.
7.37 T Give two advantages of Range Factor over fielding percentage in evalu-
ating a fielder. Do you agree with the idea that fielding percentage made more
sense back in the day when gloves were very small?
7.38 T Discuss the extent to which the plus-minus statistic depends on when
and how often a player plays in a game.
7.39 T Basketball Reference defines “true shooting percentage” (TSP) as points
divided by 2(FGA+.44FTA). Compared to effective field goal percentage (EFG),
compare how TSP and EFG handle three-point shots, missed shots, and free
throws.
7.40 T When approximating the number of possessions by a team, explain why
offensive rebounds are subtracted.
148 Sports Math

7.41 T Explain why a team’s defensive rebounds would not necessarily equal
its opponent’s field goals missed.
7.42 T Discuss why the NFL Quarterback Rating limits each of its four com-
ponents to being between 0 and 2.375.
7.43 T Explain why a rating with coefficients calibrated to accurately estimate
certain values (e.g., points) in one year might not accurately predict the same
values for the next year.
7.44 T Describe a situation in which a team’s statistics for the first half of the
season might not accurately predict its statistics for the second half of the season.
Explain why determining coefficients using odd-numbered games and testing on
even-numbered games can be better than determining coefficients with the first
half of the season and testing on the second half of the season.
7.45 T A coach praises a player, and the player’s performance declines; the
coach yells at a player, and the player’s performance improves. Explain this
phenomenon with regression to the mean. Discuss the psychological implications
of this.
7.46 T (a) After a golfer shoots a personal best score, what would you predict
happens the next round? Explain. (b) After a mediocre team gets a great upset
victory, what would you predict happens the next game? Explain.
7.47 T The coefficient for pass completion percentage is much smaller in our
regression than in the NFL Quarterback Rating formula. Compared to 2015,
in 1973 teams threw very few short passes. Explain why completions are less
valuable in 2015 than they were in 1973.
7.48 C Look up runs scored and allowed for all MLB teams in 2014. Find the
exponent (to two digits) that minimizes the sum of the squares of the errors in
Pythagorean Method predictions.
7.49 C Repeat exercise 7.48 for (a) 2014 NFL; (b) 2014-15 NBA. How do the
exponents relate to the average number of points scored in a game?
7.50 C Repeat exercise 7.48 for 2014-15 (a) NHL; (b) EPL. Decide whether
you want to predict wins (a draw is half a win) or points. Does the Pythagorean
Method seem more or less accurate for hockey and soccer compared to baseball?
7.51 C Stanley Rothman’s linear equation for NFL wins is 0.5 + .001538(PF −
PA) for a team that scores PF points and allows PA. For the 2014 season,
compute the sums of squares of errors for the Pythagorean Method with exponent
2.5 and for the Rothman method. Which method performs better?
7.52 C In this exercise, we use calculus to derive Rothman’s equation from
the Pythagorean Method. Think of the Pythagorean formula as giving wins as
2
a function of runs scored x and runs allowed y: w(x, y) = x2x+y2 . The linear
approximation of this function is L(x, y) = w(a, b) + wx (a, b) ∗ (x − a) + wy (a, b) ∗
(y − b) where wx and wy are the partial derivatives (to be explained) and a and b
are typical values of x and y, respectively. Take a = b = 725, an average number
of runs scored by a team in 162 games. The first term in the linear approximation
is w(725, 725) = 0.5 as in Rothman’s equation. Next, compute wx , which is the
derivative of w treating x as the variable and y as a constant. This is a quotient
2
rule and wx = (x22xy +y 2 )2
1
. We get wx (725, 725) = 1450 ≈ 0.0007. Show that
1
wy (725, 725) = − 1450 so our linear equation is w(x, y) ≈ 0.5 + 0.0007(x − 725) −
 Saber- and Other Metrics 149

0.0007(y −725) = 0.5+0.0007(x−y) which is a rounded off version of Rothman’s

equation.
a
7.53 C Find the linear approximation for the more general w(x, y) = xax+ya
and compare to Rothman’s equations 0.5 + .001538(P F − P A) for the NFL,
0.5 + .000351(P F − P A) for the NBA, and 0.5 + .002102(GF − GA) for the
NHL.
7.54 C Create separate scatter plots for each MLB statistic in 2004 and 2014 (in
each case, use runs scored for y) and compare, commenting on any differences.
(a) RBI; (b) HR; (c) SLG.
7.55 C Create scatter plots for each NFL statistic in 2014 (in each case, use
points scored for y) and comment on which statistic is the “best.” (a) rushing
yards; (b) passing yards; (c) rushing yards per attempt; (d) passing yards per
attempt.
7.56 C Create scatter plots for each NBA statistic in 2014 (in each case, use
points scored for y) and comment on which statistic is the “best.” (a) field goals
made; (b) field goal percentage; (c) rebounds; (d) offensive rebound percentage.
7.57 C Compute the persistence (autocorrelation) for each statistic for teams
in 2013 and 2014. (a) Football yards rushing; (b) Football yards passing; (c)
Basketball field goal percentage; (d) Basketball free throw percentage.
7.58 C For NBA teams in 2014-15, construct scatter plots and compute corre-
lations using points as one variable with the other variable being (a) FTM/FGA;
(b) FTA/possessions. (Estimate possessions using FGA−OREB+0.4FTA+TO.)
Discuss the use of FTM/FGA in the Four Factors.
7.59 C Compute the average NFL Quarterback Rating for (a) 1973, (b) 1983,
(c) 1993, (d) 2003, and (e) 2013. Discuss how the game is changing.
7.60 C For the English Premier League in 2013/14, find the best linear equation
for team goals in terms of shots, possession time, passing percentage, and fouls.
(Data can be found at whoscored.com.) Discuss the influence of each variable.
7.61 C Find the best fit line for the data (1,1), (2,4), (3,9). Explain why it
is silly to fit a line to this data. Given this example, discuss the importance of
graphing your data before finding a regression equation.
7.62 P Track the errors in the Pythagorean Method for some sport. Compute
the autocorrelations for consecutive years. (The correlation of both lists, one of
which has the luck values for each team in one year and the other of which has
the luck for the teams in the same order the next year.) Does “luck” seem to be
the right word for the errors?
7.63 P Taking pace of play into account, study whether the better offensive
team in a championship game/series wins more often than the better defensive
team wins.
7.64 P The Four Factors are considered to be “orthogonal” so that the corre-
lation between any two of them should be near zero. Use team season totals to
explore whether the factors are orthogonal.
7.65 P Find the value of k that best predicts runs scored by a team in a season
with the formula k OBP + SLG. How does OPS (k = 1)
7.66 P For your favorite sport, compute the “winning percentage” for a variety
150 Sports Math

of statistics. For example, in football you could look at TD passes; find the
percentage of games in which the team with more TD passes wins the game.

Further Reading
The book Moneyball is by Michael Lewis, who also helped with the screen-
play for the movie. Big Data Baseball is a more recent description of the
Pittsburgh Pirates’ conversion to analytics and their rise to prominence. The
Only Rule Is It Has to Work describes a season using analytics to run a minor
league team.
The Sloan Sports Analytics Conference baseball panels (videos available
online) have much more information about Oakland’s scouting practices. Nate
Silver’s excellent book The Signal and the Noise also mentions Oakland’s
scouting budget.
A derivation of the general Pythagorean Method is given in Steven J.
Miller’s “A derivation of the Pythagorean Won-Loss Formula in Baseball.”
The Music of Pythagoras is an enjoyable review of what is known and
not known about the historical Pythagoras and his followers. The Cult of
Pythagoras takes a more skeptical look at the evidence.
The 1982-88 Bill James Baseball Abstracts were published by Ballantine
Books, and are enjoyable reads even today. Hardball Times and Baseball
Prospectus publish baseball annuals that are in the same vein that James
mined.
Annual guides include The Fielding Bible from Baseball Info Systems, the
Football Outsiders Almanac, Hardball Times’ Baseball Annual, and guides
from the Prospectus family (Baseball, Basketball, and Hockey).
The Numbers Game: Why Everything You Know About Soccer is Wrong
discusses scoring trends in various soccer leagues. A summary can be found
at the soccerbythenumbers.com web site. Soccermatics is a nice complement,
exploring other forms of analytics.
Excellent books about sports analytics include Baseball Between the Num-
bers by Baseball Prospectus, The Sabermetric Revolution by Baumer and
Zimbalist, Mathletics by Winston, Basketball on Paper by Oliver, Basketball
Analytics Spatial Tracking by Shea, Analytic Methods in Sports by Severini,
Analyzing Wimbledon by Klaasen and Magnus, Stumbling on Wins by Berri
and Schmidt, and The Book by Tango and Lichtman.
Chapter 8
Randomness in Sports

Introduction
The lead pass to Marvey’o
Otey was thrown too far. Otey,
playing for William Byrd High
School in Vinton, Virginia, on
December 9, 2013, chased the
ball down but only had enough
time to get his right hand on the
ball and whip it behind his back.
Then Otey’s momentum carried
him out of bounds and through FIGURE 8.1: Behind the Back and
an open door out of the gym and Straight Out the Door
into a school hallway. Hearing
the crowd cheering, Otey thought his desperate save must have been grabbed
by a teammate. Only later did he learn that his save had gone in the basket
for one of the most outrageous three-point baskets ever.
There is no doubt that luck plays a role in sports. Otey’s shot, the football
pass that ricochets off of three defenders right into a receiver’s hand for a
touchdown, the line drive that bounces off the wall at an odd angle: these
pieces of good luck are balanced by the shot that beats the goalie only to
bounce off of a post, or the putt that hits the wrong blade of grass and veers
away from the hole.
The extent to which luck affects results is open to debate, and is the subject
of this chapter. As we explore this general issue, the following questions will be
addressed. Which sport is most subject to chance? Does the best team always
win? Which sports leagues have the most balance? Does the “hot hand” exist?
Do balanced scoring teams win the NBA championship? Are base hits a matter
of luck?

151
152 Sports Math

Summing Up the Basics

Some basic probability and statistics tools are needed to follow many ana-
lytics discussions, whether they be about sports or politics or business. Mean,
standard deviation, and distributions are briefly introduced here. To make the
discussion more concrete, we use the two sets of numbers g = {27, 34, 22, 21,
30, 28} and r = {17, 12, 31, 18, 9, 15}. These are the points scored in the first
six games of the second round of the 2015 NBA playoffs by Blake Griffin and
J.J. Redick, respectively.
To see which player is higher scoring, you can compute the means. The
mean is what most people think of as “the average” even though other ver-
sions of averages (like the median) are preferred by statisticians. The mean
of g is 27+34+22+21+30+28
6 = 27 and the mean of r is 17+12+31+18+9+15
6 = 17.
Griffin is higher scoring.
Both players showed several ups and downs in their scoring patterns, but
Griffin was much more consistent than Redick. We measure consistency with
variance and standard deviation. To compute the variance, usually denoted
s2 , subtract the mean from each value, square the result, add and divide by
one less than the number of values (6 − 1 = 5). For g, we first compute {0,
7, −5, −6, 3, 1}, then {0, 49, 25, 36, 9, 1}, then the sum 120 and finally the
variance 120 290
5 = 24. For r, we get a variance of 5 = 58.
The standard deviation is simply the square root of the variance, and
is denoted by s. (We denote the variance by s2 so that the standard deviation
has a simple representation. This is an indication that the standard√deviation
will be more useful to√us than the variance.) For Griffin, we get s = 24 = 4.9
and for Redick s = 58 = 7.6. This quantifies the fact that Griffin’s points
were more consistent.
The empirical rule of statistics states that for bell-shaped data, about
68% of the points will be within one standard deviation of the mean, 95%
within two standard deviations of the mean, and 99.7% within three standard
deviations of the mean. We have no information about whether or not an
individual player’s point totals are bell-shaped, but let’s see how this might
work. For Blake Griffin, within one standard deviation of the mean gives the
range 27±4.9, or between 22 and 32; in fact, 4 out of 6 (67%) of his point
totals are in this range. Plus or minus two standard deviations is 27±9.8, or
from 17 to 37; all of his point totals are in this range. For Redick, two-thirds of
the point totals are in the range 17±7.6, or between 10 and 24. Two standard
deviations gives the range 17±15.2, or between 2 and 32; this contains all
of his values. Notice that with Redick’s inconsistency a much wider range of
values is needed to capture his actual output.
The empirical rule assumes a bell shape for the data. The distribution
of data is often overlooked in basic analyses. The bell curve of the normal
 Randomness in Sports 153

distribution is a common occurrence, but it should be checked before using

the empirical rule.
Figure 8.2 is a histogram or bar
graph of points scored by the Los
Angeles Clippers during the 2014-15
NBA regular season. Each bar rep-
resents a point range of four points,
and is drawn to a height represent-
ing the number of times data points
in that range occurred. For exam-
ple, the tallest bar shows that there
were 16 games in which the Clippers FIGURE 8.2: Histogram
scored between 104 and 108 points. The bars do not form a perfect bell curve,
but there is a clear peak in the middle with a nearly symmetric drop-off to
each side.
Figure 8.3 overlays a bell curve
on top of the histogram of Figure 8.2.
(The equation for this curve will be
explored in the exercises.) With this
visual, we can see that the histogram
is approximated reasonably well by a
bell curve. This is enough to justify
using the empirical rule. Even an 82-
game season is not long enough for
us to demand that the distribution FIGURE 8.3: Normal Curve
form a perfect bell curve.
There are numerous distributions
of importance that are not normal.
A course in probability will intro-
duce several common distributions.
To illustrate a different distribution
that occurs with regularity, Figure
8.4 shows a histogram for points per
game in 2014-15 for all NBA play-
ers who averaged at least 2 points
per game. The first bar from the left
shows that 82 NBA players averaged FIGURE 8.4: NBA Scorers
between 2 and 5 points per game in
2014-15. As the scoring average increases, the number of players with that
average decreases at a regular rate. This is an example of a power law distri-
bution (to be explored in exercise 8.49).
154 Sports Math

Prediction is Difficult
The physicist Niels Bohr is credited with saying, “Prediction is very dif-
ficult, especially about the future.” It is not recorded whether or not he had
just visited Las Vegas. Most sports bets are “against the spread” so that the
bet feels fair. You may have no doubt that San Antonio will defeat New York,
but if the spread is 15 points then San Antonio must win by more than 15
points for you to win a bet on San Antonio.
To illustrate how risky such a bet is, suppose that your friend is a big New
York fan, and is willing to bet on New York with a spread of 3 points. The
experts who determine the spread are saying that San Antonio is 15 points
better, but the Spurs need to win by only 4 points or more for you to win
your bet. That sounds like a great bet! But, how likely are you to win? Take
a guess: 80%? 90%? higher?
To answer the question, we need to know the distribution of NBA scores.
It turns out that the difference between the spread and the actual outcome
is (approximately) normally distributed with mean 0 and standard deviation
12. On average, then, the spread gives the outcome: San Antonio by 15. Using
the empirical rule, we know that about 68% of the results will be within one
standard deviation of the mean. So, 68% of the results will have San Antonio
winning by between 15 − 12 = 3 and 15 + 12 = 27 points. More importantly,
the other 32% of the time the result will not be between 3 and 27, with 16%
less than 3 and 16% greater than 27. You will lose your bet 16% of the time!
It’s true that having an 84% chance of winning a bet is good, but this is
probably not as high a probability as you were expecting.
Wayne Winston’s book Mathletics gives standard deviations for scores in
several leagues. The NBA is 12 points, and college basketball is 10 points; I
would have guessed that the younger college players would have more fluctu-
ations in their scores. The standard deviation for NFL scores is 14 points; for
college football scores it is 16 points. Even though the total points scored is
lower, scores in football vary more than in basketball!
Suppose that you have a system that consistently beats the NFL point
spread by 4 points. What percentage of bets will you win? Take a guess, but
keep in mind that the NFL standard deviation is 14 points. Calculus can be
used to compute that you should win a little more than 61% of the time. If you
want to win 65% of your bets, you need to increase your margin of superiority
over the spread to 5.4 points.
It is not easy to have a high winning percentage in Las Vegas!
 Randomness in Sports 155

A Slump or a Disaster
Two main goals of analytics are prediction and evaluation. We have seen
that randomness can make prediction difficult. The same is true of evaluation.
Suppose that a baseball team opens the season with 10 wins and 15 losses. Is
this a bad team? If the team has a lot of talent, is it time to fire the manager?
A baseball team with a winning percentage of 60% will win 97 games over
a 162-game season. This is more than enough wins to make the playoffs. Is
it possible that such a team could have a 10-15 start just by bad luck? The
question is asking us to compute the probability that a 60%-quality team could
have a 10-15 start. To answer this question, we start by making assumptions,
giving us a model with which probabilities can be computed.
The simplest model is a binomial model under which games are
Bernoulli trials. The terminology is important because it gives us a short-
hand to describe a common set of assumptions. In particular, a game has two
possible outcomes (win or lose), and we assume that our team has a 60%
chance of winning each game. That means that the games are independent:
the outcomes of past games do not affect the probability of the next game. We
are thus ignoring the effects of good and bad pitching opponents (every game
has the same win probability) as well as the effects of good and bad streaks
(games are independent). We do not have to believe that this is true, but it
allows us to compute a probability and draw conclusions.
We can evaluate the probability in multiple ways. We start by calculating
mean and standard deviation. To do so, we use the following facts about
binomial distributions. If x is the number of times a particular outcome (e.g.,
a win) of a sequence of n Bernoulli trials of probability
p p occurs, then the
mean of x is np, and the standard deviation for x is np(1 − p). The following
example shows how this works.

Example 8.1 If a baseball team wins games with probability p = 0.6 and
the games are independent, compute the mean and standard deviation of the
number of wins in 25 games.
Solution The main assumptions of a binomial distribution are present: two
outcomes (win or lose) per try, constant probability, and independence. We
have probability p = 0.6 and n = 25 trials. The mean is np = 25·0.6 = 15. This
should make sense: 60% of 25 is 15. The standardp deviation is√not as common
sensible,
√ so it is nice to have the formula σ = np(1 − p) = 25 · 0.6 · 0.4 =
6 ≈ 2.45.

The empirical rule applies for binomial distributions with large values of
n. We expect the number of wins to be within two standard deviations of the
mean 95% of the time. In this case, we look at 15 ± 2 · 2.45 or 10.1 to 19.9. Our
team’s 10-win total is right at the edge of this. What do we conclude? If the
156 Sports Math

games were completely random, a bad streak worse than 10-15 would occur
about 2.5% (half of 5%) of the time, or about one in 40. Purely on the basis
of the team’s record, there is reason to stay patient: bad streaks can occur by
chance. Of course, the games are not completely random. If the 25 games were
all at home against bad teams and the manager was alienating players and
fans, then perhaps a change of manager is in order. However, keep in mind
regression to the mean, which says that this team is likely to bounce back to
its 60%-win mean.

Calculus Box: Probability

The probability in Example 8.1 can be computed directly. Let’s start with
a smaller example. What is the probability that the 60% team starts with 1
win and 3 losses? For example, the team could win its first game and then
lose three in a row. The game-by-game probabilities of this happening are 0.6,
0.4, 0.4, and 0.4. Because the games are assumed to be independent events,
we can multiply these probabilities together to get a probability of (0.6)(0.4)3
for the sequence WLLL. This is not the only way to start 1-3. Notice that
the sequence LWLL would generate the same probability: 0.6 · 0.4 · 0.4 · 0.4 =
0.4 · 0.6 · 0.4 · 0.4 = (0.6)(0.4)3 . So, the total probability of starting 1-3 will
equal (0.6)(0.4)3 times the number of different orderings for one W and three
L’s. We have listed WLLL and LWLL so far; how many more are there? One
way to answer this is to focus on the one win: there are four games in which
that win could occur, so there are four different orders. The probability is
4(0.6)(0.4)3 .
The general structure of our answer can be used for all problems of this
type. If there are W wins and L losses, then
Binomial probability = (Number of orders) * pW * (1 − p)L .
All that remains is to find a simple way to count the number of se-
quences. This is done using a formula called the binomial coefficient. For
our calculation, we want to know how many different ways 10 wins can
be arranged in a sequence of 25 games. The binomial coefficient
for this is
named “Twenty-five choose ten” and has several notations: 25 10 , 25 C10 , and
C(25,10) are common ones. The formula uses the factorial which is defined
as n! = n(n − 1)(n − 2) · · · (2)(1). For example, 5!=5 · 4 · 3 · 2 · 1 = 120. Then
the binomial coefficient equals
 
n n! n(n − 1)(n − 2) · · · (n − m + 1)
= =
m m!(n − m)! m(m − 1)(m − 2) · · · 1
where the second version of the formula emphasizes that the same number of
factors are in the numerator and denominator. Before tackling the numbers
 Randomness in Sports 157

in Example 8.1, a simple calculation shows that 63 = 3·2·1

6·5·4
= 20. This is the
number of different orders of wins and losses with 3 wins in 6 games.

Example 8.2 If a baseball team wins games with probability p = 0.6 and
the games are independent, compute (a) the probability of winning exactly 10
games out of 25; (b) the probability of winning at most 10 games out of 25.
Solution The assumptions of the binomial distribution are met.
(a) The prob-
ability of winning 10 games and losing 15 is given by 25 10 (.6)10
(.4)15
with
25 25! 25!

10 15
10 = 10!15! . Putting this all together, the probability is 10!15! (.6) (.4) ≈
0.0212. (b) To the value in part (a), we need to add the probability of winning 9
25! 25!
games, 9!16! (.6)9 (.4)16 , and the probability of winning 8 games, 8!17! (.6)8 (.4)17 ,
and so on. The sum can be written as
n=10
X 25!
(.6)n (.4)25−n
n=0
n!(25 − n)!
and equals approximately 0.0344.

You may have noticed that the value in Example 8.2 is larger than that of
Example 8.1. This is a result of applying the empirical rule when the number
of games is not large. The issue has to do with the empirical rule applying
to continuous distributions in which the probability of a single outcome is
essentially zero. There is no distinction between the probability of the value
of the variable being less than 10 versus being less than or equal to 10. In
Example 8.2, the probability of the number of wins being less than or equal
to 10 is 0.0344, but the probability of wins being less than 10 is 0.0132. The
average of the two is 0.0238, which is very close to that given by the empirical
rule.

May the Best Team Win

A playoff-caliber team can have a slump over a 25-game stretch. With a
162-game schedule, it seems likely that the best baseball team will prevail in
the long run. A quick calculation puts that in doubt, and a simulation run by
Bill James gives surprising results.

Example 8.3 If a baseball team wins games with probability p = 0.6 and
the games are independent, compute the mean and standard deviation of wins
over a 162-game season. If it takes 91 wins to make the playoffs, how likely is
the team to do so?
Solution The assumptions of the binomial distribution are met.
p The mean is
(0.6)162 = 97.2 or about 97 wins. The standard deviation is (162)(.6)(.4) =
158 Sports Math

6.2 or about 6 wins. Since 91 wins is one standard deviation away, the empirical
rule gives the team a 16% chance of having less than 91 wins, or an 84% chance
of making the playoffs. (The calculus box above shows how to get the exact
value, which is about 86%.)

The 1989 Baseball Abstract includes an essay by Bill James on this issue.
James simulated 1000 seasons of the then-current league of 26 teams in 4
divisions. He first randomly assigned each team a “true quality” based on the
historical distribution of winning percentages in baseball. So, a team’s quality
was chosen randomly from a normal distribution with mean 0.5 and standard
deviation 0.05. The computer then simulated every game for a season, using
the official schedules. This raises a significant question: if team A of quality
0.55 plays team B of quality 0.52, how often should team A win? The answer
should logically be greater than 0.5 (team A is better) and less than 0.55
(team B is above average). James used the conditional probability formula
0.55(1−0.52)
0.55(1−0.52)+(1−0.55)0.52 = 0.530, which is now known as the log5 method.
The results of the simulation are surprising. The best team (greatest true
quality) in a division won the division 54.6% of the time - barely over half!
The best team in baseball failed to win its division 28.5% of the time, and
won the simulated World Series 29.3% of the time. That is, the best team was
as likely to miss the playoffs as it was to win it all.
A problem that James had with his simulation helps explain the surpris-
ing result. The actual spread of wins and losses in the simulations produced
a larger standard deviation than the value that he used to create the true
qualities. This is due to the number of wins for one team not really being
independent of the number of wins for another team. If an underdog team A
gets an unexpected (“lucky”) win over division rival team B, then team B also
just got an unexpected loss. A small number of upsets between division rivals
can dramatically change the standings since division games count double (a
win for one plus a loss for the other).

Measuring Parity: Gini in a Bottle

Which of the major sports leagues has the most parity? Leagues such as
the NFL enforce salary caps and arrange the draft of new talent to favor the
teams with the worst records. Does this actually result in parity? In the first
49 years of the Super Bowl, 19 of the 32 teams (57%) won at least once while
28 (87.5%) made the Super Bowl at least once. In the last 49 major league
baseball seasons, 20 of the 30 teams (66.7%) won at least one World Series
and 27 of the 30 teams (90%) made it to the Series at least once. Baseball,
in spite of a limited salary cap and decreased importance of the draft, shows
 Randomness in Sports 159

more parity. In the last 49 years of the NBA, 15 out of 30 teams (50%) won
the title while 21 out of 30 (70%) made the finals. There seems to be less
parity in basketball.
There are ways to measure parity
other than simply counting up champions
and runners-up. The Gini index is used
by economists to measure inequality of
wealth. Imagine a league with the ultimate
parity: each of 100 teams wins exactly half
of its games, with 5 wins and 5 losses. Take
the bottom 5% of the teams (the five worst
teams) and add up their wins. We have 5 FIGURE 8.5: Lorenz Equality
teams with 25 wins, which is exactly 5% of
the 500 total wins in the league. The bottom 10% of the league (10 teams) will
have 50 total wins, exactly 10% of the total wins in the league. The Lorenz
curve of percentile versus percentage of wins would include the points (5,5)
and (10,10) based on our discussion so far. The full graph would be the straight
line shown in Figure 8.5.
Now, imagine the same league with a
large disparity, with the five worst teams
each losing all 10 games and the next five
worst teams winning twice each. The total
number of wins for the bottom 5% is 0, so
the Lorenz curve includes the point (5,0).
The bottom 10% of the league won a total
of 10 games, which is 2% of the 500 total
wins. The Lorenz curve includes the point FIGURE 8.6: Lorenz Unequal
(10,2). If the next 10 teams each won 3
games, then the bottom 20% of the league won 10 + 30 = 40 games, which
is 8% of the total: the Lorenz curve includes the point (20,8). The full graph
might look like Figure 8.6.
The more inequality there is in the
league, the farther the Lorenz curve will get
from the total-equality curve of Figure 8.5.
So, the gap between the Lorenz curves in
Figure 8.5 and 8.6 is a measure of the dispar-
ity in the league. The Gini index is a ratio
of areas that quantifies this statement. The
denominator is the area between the ideal
curve (Figure 8.5) and the x-axis. This is a FIGURE 8.7: Lorenz Curves
triangle of width 100 and height 100, which
has area 5000. The numerator is the area between the actual Lorenz curve
and the ideal curve, as shown in Figures 8.7 and 8.8.
160 Sports Math

The calculation of the area for regions

such as the one shown in Figure 8.8 can
require calculus. An estimate can be made
from the data by summing up areas of
trapezoids. Suppose that the data points
shown in Figure 8.6 are (0,0), (5,0), (10,2),
(20,8), (40,24), (60,44), (80,68), (90,82), and
(100,100). The area of a trapezoid is the base FIGURE 8.8: Gini Area
times the average of the heights. For the re-
gion using the x-axis and the first two data points (0,0) and (5,0), the width
is 5 − 0 = 5 and the average height is 0+0 2 = 0. The next region has width
10−5 = 5 and average height 0+2 2 = 1. The third region has width 20−10 = 10
and average height 2+8 2 = 5. The area under the Lorenz curve is given by
5 · 0 + 5 · 1 + 10 · 5 + 20 · 16 + 20 · 34 + 20 · 56 + 10 · 75 + 10 · 91 = 3835. The area
between the curves is the difference 5000 − 3835 = 1165 and the Gini index is
1165
= 0.233.
5000
Example 8.4 A basketball team has players score 34, 28, 26, 8, and 4 points
in a game. Plot the Lorenz curve and calculate the Gini index.
Solution The total points scored by the team is 100. The bottom 20% (the
lowest of five) scored 4% of the points. The bottom 40% (the two lowest
scorers) scored 12% of the points, and so on. The points for the Lorenz curve
are (0,0), (20,4), (40,12), (60,38), (80,66), and (100,100). The Lorenz curve
that connects these points with line segments is shown in Figure 8.9. As before,
the denominator for the Gini index is 5000. The numerator is the difference
between 5000 and the sum of the areas of trapezoids. The trapezoid areas are
given by width times average height. We have 20 0+4 2 + 20 2
4+12
+ 20 12+38
2 +
38+66 66+100 5000−3400
20 2 + 20 2 = 3400. The Gini index equals 5000 = 0.32.

The Gini index is always between 0 and

1, with 0 representing perfect equality and
1 representing total inequality. In Example
8.4, since we are subtracting from 5000, the
maximum numerator is 5000, which would
produce a Gini index of 1. To subtract 0,
the points need to be (0,0), (20,0), and so
on, with the last player scoring 100% of the
points. The smallest possible numerator is 0, FIGURE 8.9: For Ex. 8.4
which occurs if we subtract 5000. This hap-
pens if the Lorenz curve is y = x, occurring if every player scores the same
number of points.
Gini indices for 2014 are shown in Table 8.1. By this measure, then, base-
ball has the most equality. This should make sense, given that there are often
NBA and NFL teams that win 80% of their games and the best baseball team
barely wins 60% of its games.
 Randomness in Sports 161
TABLE 8.1: Gini Indices
NBA 0.18
MLB 0.06
NFL 0.17

Measuring Parity: Luck versus Skill

A different way to measure league balance can be thought of as measur-
ing the percentage of luck in the sport’s games. This terminology is a little
misleading, but the idea is to look at win variance in a league compared to a
league whose games are completely random. The 2014-15 NBA season gives
us a test case.
Imagine two alternate-universe NBAs, one in which every game is deter-
mined by a coin flip and one in which the better team always wins. Call these
the Luck-NBA and the Skill-NBA. We can compute the variance of wins for
each league. The Luck-NBA has a binomial distribution with n = 82 games
and p = .5 for each game. The variance for wins of an individual team is
np(1 − p) = 20.5. In the Skill-NBA, the best team goes 82-0, the second-best
team goes 80-2 or 79-3 or 78-4 depending on how many times the teams played
(determined by which divisions the teams are in). Using a value of 3 games per
team, the variance of the number of wins is about 620. The variance for the
number of wins in the real NBA in 2014-15 was about 181. The real variance
is between the Skill-NBA variance and the Luck-NBA variance. Our goal is
to determine the right combination of skill and luck to produce the actual
variance.
The model we use is that a game result is a fraction p of skill and a fraction
1−p of luck. If p = .4, then the result is 40% skill and 60% luck. In an equation,
we assume that
r = p s + (1 − p) k
where the luck k has mean 0 and is independent of skill. With these assump-
tions, the result variance equals p2 times the skill variance plus (1 − p)2 times
the luck variance. For the NBA, we want

181 = p2 (620) + (1 − p)2 (20.5)

which we can solve to get p = 0.53. We should not conclude that NBA games
are half luck, but we can use this number to see how balanced the NBA is
compared to other leagues. The larger the value of p, the more often the better
team wins, and the less balanced the league is. The p value for MLB wins in
2014 is p = 0.16. Certainly, the better team wins less often in baseball than
basketball. Wins in the 2014 NFL have p = 0.61, the largest of the three. By
this measure, then, the NFL has the least balance, and the better team wins
the most often.
162 Sports Math

The Paradox of Skill

A thought experiment takes this in a
different direction. Suppose that all players
have the same high skill level. Then every
contest would be determined by a gust of
wind, bounce of the ball, or referee’s call; in
other words, luck. This is the paradox of
skill: the less variation in skill there is, the
more important luck is. Now, think of the
skill levels of players being described by a FIGURE 8.10: Skill Curve
bell-shaped curve. Most players are average,
but a few are especially good and a few are especially bad. Figure 8.10 shows
a bell curve with a vertical line drawn to the right to indicate a maximum
human level of performance.
If all players get better, what happens?
At first, the curve in Figure 8.10 could sim-
ply move to the right. However, as the dis-
tribution begins to approach the limit of hu-
man performance, the distribution will nec-
essarily pile up near the limit. Figure 8.11
shows the beginning of this process, with the
obvious side effect of the spread of the dis-
tribution decreasing.
Stephen Jay Gould used this idea to ex- FIGURE 8.11: More Skill
plain why no baseball player has had a season batting average over .400 since
Ted Williams in 1941. In baseball, pitchers and rule makers combine to keep
league batting averages from increasing. A league-wide increase in baseball
skill would not result in increased batting averages. By the above argument,
the result could be a reduced standard deviation, which would reduce the odds
of someone having a remarkably high batting average. To see why, consider
two Boston Red Sox hitters, Ted Williams batting .406 in 1941 and Wade
Boggs batting .368 in 1985. Which is more impressive? In Keep Your Eye on
the Ball, Watts and Bahill compute league averages of .282 in 1941 and .268
in 1985, and standard deviations of .0340 in 1941 and .0264 (smaller!) in 1985.
Williams’s average is 3.65 standard deviations above the league average, and
Boggs’s average is 3.81 standard deviations above the league average. Boggs’s
.368 average in 1985 is more out of the ordinary than Williams’s .406 average
in 1941, due to 1941’s higher league average and higher standard deviation.
As athletes get better, it becomes harder to be clearly superior to all others.
To put this in terms of the paradox of skill, the better the competition is the
more you need luck (and lots of it) to dominate.
 Randomness in Sports 163

Measuring Parity: Entropy

A third way to measure league balance is to adapt the concept of entropy.
A cornerstone of information theory, entropy (or Shannon entropy, named for
Claude Shannon) measures the amount of information in a signal in the sense
that the more predictable a message is, the less information is gained when
the message arrives. Maximum entropy occurs when all possible signals are
equally likely (so that the signal is as unpredictable as possible). In a sports
context, entropy will be maximized when every team or player has an equally
likely chance to win or score; in other words, when there is maximum parity.
If there are n possible outcomes which occur with probability p1 , p2 , . . . , and
pn then entropy is defined by

entropy = −p1 ln(p1 ) − p2 ln(p2 ) · · · − pn ln(pn )

where “ln” is the natural logarithm.

Example 8.5 The eight highest scorers in points per game in the playoffs
leading up to the 2015 NBA Finals were c = {27.6, 18.7, 13.5, 10.1, 9.4, 9.1,
7.0, 4.8} for Cleveland and g = {29.2, 19.7, 14.0, 11.3, 8.0, 5.3, 5.0, 4.9} for
Golden State. (Data from basketball-reference.com.) Compute the entropy for
each team.
Solution For Golden State, the total points for the eight players is 97.4, so
divide each value in g by 97.4 to get the proportion of points gp = {.300, .202,
.144, .116, .082, .054, .051, .050}. For Cleveland, the total points for the eight
players is 100.2, so the proportion of points scored is essentially equal to the
values in c divided by 100. The entropy for Cleveland is
−.276ln(.276) − .187ln(.187) − .135ln(.135) − .101ln(.101) − .094ln(.094) −
.091ln(.091) − .07ln(.07) − .048ln(.048) = 1.94.
By a similar calculation, the entropy for Golden State is 1.88.

For comparison, the maximum entropy for eight scorers is 2.08 (ln8), and
a team for which one player scored 99.3 percent of the points would have
entropy 0.05. (Technically, ln(0) is undefined so a one-person team would
have undefined entropy, but this shows that the limiting entropy value for
complete inequality is zero.) The calculation shows that Cleveland was more
balanced in its scoring than Golden State.
In the 2015 playoffs, the team with the higher scoring entropy in the regular
season (top ten scorers) won just 5 of the 15 playoff series. The NBA is a league
of stars.
A similar calculation can be made for the entropy of wins in a league. Table
8.2 shows the entropy values for 2014 in the middle column. The entropy for
baseball is higher than for basketball, indicating again that baseball has more
parity. The NFL entropy is almost as high as the MLB value, but that is
164 Sports Math

misleading since the NFL has more teams. The right-most column shows the
league entropy divided by maximum entropy. We can see that baseball has,
by far, the most parity while the NBA has more parity than the NFL.
TABLE 8.2: League Entropy
entropy pct of max
NBA 3.34 0.984
MLB 3.39 0.998
NFL 3.38 0.975

Declaration of Independence
In several examples, we have assumed that events such as games are in-
dependent processes. This makes the calculations easier, but we should worry
about how unrealistic the assumption is. In Analyzing Wimbledon, Klaasen
and Magnus conclude that points in a tennis game are won in an independent-
like pattern. That is, if the server wins a point on serve 57% of the time, cal-
culations of games won, deuces won, and so on, using a binomial model with
p = 0.57 match the actual results reasonably well.
For the 2014 NBA Finals, the NBA switched from a 2-3-2 scheduling of
home games to a 2-2-1-1-1 schedule. In 2013 the first two games were in Miami,
the next three in San Antonio, and the last two in Miami. In 2014, the first
two games were in San Antonio, the next two in Miami, and the fifth game in
San Antonio, with the sixth game scheduled in Miami and the seventh game
scheduled in San Antonio. If the games are independent, the sequence does
not matter: the first four and the first six games are split equally in both
scenarios, and the independence assumption means that order does not affect
the probabilities (more on this in the exercises).
To check for independence, we can look at a variety of situations. In the
NBA playoffs from 2003 to 2015, the home team won 65% of the games.
If games are independent, this percentage should carry through all games.
However, the home team won game one 72% of the time, game two 75% of
the time, game three 56% of the time, and game four 54% of the time. There
is an explanation for this difference. Games one and two are played at the
home court of the better team, games three and four at the home of the lesser
team. We can claim independence as long as we assign different probabilities
for the two teams: 75% for the better team and 55% for the worse team seem
reasonable.
Let’s take a closer look at game five. The independence assumption would
imply that the home team wins 75%, very close to the actual 78% figure.
Independence looks good so far! The home team in game five could be ahead
 Randomness in Sports 165

3 wins to 1, or tied 2 wins to 2, or behind 1 win to 3. In 3-1 situations, the

home team won 76% of the games; in 2-2 situations, the home team won 75%.
However, in 1-3 situations (where another loss would eliminate the team from
the playoffs) the home team won 27 out of 31 times, or 87%! To see if this
is statistically significant, we can compute the probability that a 75% team
would randomly win 27 or more games out of 31: the binomial model gives an
8% chance, so this does not quite qualify as statistically significant. But it is
suggestive that the independence assumption does not apply to a team facing
elimination!

Conditional Probability
The calculations done above are examples of conditional probabilities.
Instead of looking at how often the home team wins game five, we look at how
often the home team wins game five given that the series is tied 2-2 or given
that the team trails 1-3. The probability could be changed by the condition
that is imposed (e.g., game five with the series tied).
A formal definition of independence is that events A and B are independent
if the conditional probability of event A given that event B occurs equals the
(unconditional) probability of A. That is, A has the same probability whether
or not B occurs. In our NBA playoff situation, if the outcome of game five
is independent of the current won-lost standings, then the probability of the
home team winning is the same whether it is ahead 3-1, tied 2-2, or behind
1-3. As noted, there is some evidence that this might not be true.
When we say that game outcomes are independent, or the outcomes of at
bats or free throws or whatever are independent, we are really saying that
they are independent of everything. The probability of success is not changed
by anything (the current score, the loud fan in the second row, the current
alignment of Jupiter and Mars, anything). The challenge for analysts is to
actively test for factors that could change the probability. Otherwise, the
analyst’s calculations could be meaningless.
A silly, but (unfortunately) common, example of conditional probability
is the habit of announcers to give situational statistics. Bob Uecker “played”
baseball before his excellent announcing and acting career, and in true Uecker
style managed to finish his career with a batting average of .200. However,
he batted .300 against Hall-of-Famer Steve Carlton and .333 against Hall-of-
Famer Warren Spahn! Before you get excited, notice how suspiciously round
those averages are. In fact, Uecker was 2-for-6 against Spahn and 3-for-10
against Carlton. The numbers are not large enough to be significant: for ex-
ample, a .200 hitter would randomly get 2 or more hits in 6 at bats more than
one-third of the time. (You can find data like this at baseball-reference.com.)
A better example, although only slightly better, involves Pete Rose. Rose’s
166 Sports Math

first hit was off of Bob Friend, against whom Rose was a robust 16-for-36. His
last hit was off of Greg Minton, an excellent pitcher whom Rose touched for
13-out-of-30. As you will see in exercise 8.60, even 30 or 36 at bats is not
enough to draw much of a conclusion.

The Hot Hands

All athletes know the feeling of being “in the zone” or “on fire.” The target
looks twice as large as normal, time moves slowly, distractions disappear, and
success seems inevitable. The widespread knowledge of the sensation of the
“hot hand” made it especially galling when two psychiatrists, Amos Tversky
and Thomas Gilovich, published research in the 1980s claiming that the hot
hand is merely a cognitive illusion. Their work centered on a simple test of
independence. If a person gets hot, the probability of success increases. If
successes are independent, then there is no such thing as the hot hand.
Tversky and Gilovich recorded all field goal attempts by Philadelphia 76er
players in home games in the 1980-81 season. They looked at the percentage
of shots made after strings of makes and strings of misses. If all shots are inde-
pendent, the percentages should be equal. If hot hands exists, the percentages
following strings of made shots should be higher than those following strings
of misses. However, they were basically the same. They looked at free throws
by Boston Celtics players during the 1980-81 and 1981-82 seasons in the same
way. Again, they found no significant differences in shooting percentages. In
fact, most players shot slightly better after a miss.
Numerous studies have followed for other sports, and the findings have
been consistently negative. Streaks do not seem to occur in professional sports
beyond what you would expect from a random sequence. It is important to
note that the probability p of success is not always 0.5. A 90% free throw
shooter will have long streaks of made free throws. The streaks just are not
significantly longer than what you would get from a coin that was biased to
come up heads 90% of the time. Few violations of independence have been
found.
The cognitive illusion aspect of Tversky-Gilovich is a flaw of the human
brain that has been demonstrated numerous times. Try this quiz: one of the
three sequences of Hs and Ts was created using a coin flip model where each
flip is independent. The other two sequences were created with processes that
are not independent. (Of course, any of the three sequences could result from
coin flipping. Randomness can be so random.)

THTTTTHTTTTHHTTHHHTHTHHTHHHTHH
HTTTHTHHTHTTHHHTHTHTTTHTHHTHTH
HTHHHHHHTTTTTTTTHHTHHHHHHTTTTT
 Randomness in Sports 167

The third sequence should look too streaky to be from coin flipping; in
fact, it was created so that the previous outcome would be repeated 80% of
the time. Do you think the first or second sequence shows independence? The
first sequence has an early streak of 8-out-of-9 Ts; the second sequence has
a nice balance of Hs and Ts in every subsequence of nine symbols. And that
is what is fake about it: the rule was to do coin flipping until three in a row
occurs, but never allow four in a row to occur. The third sequence is streakier
than real coin flipping, while the second sequence is less streaky than real coin
flipping. And yet, most people choose the second sequence as the real one, and
if asked to generate a “random” sequence of Hs and Ts most people produce
something like the second sequence.
The lesson is that the human brain is wired to find and explain patterns,
and we have a tendency to assign meaning (like “hot” or “cold”) to sequences
that are actually random.

Not So Fast, My Friend

The false lead about the first sequence above (“an early streak of 8-out-
of-9 Ts”) leads to an important critique of much of the research into the hot
hands. Making 8 out of 9 probably qualifies as hot even though it may not have
an especially long streak of consecutive makes. Depending on the situation, 7
out of 9 might be noteworthy. Research that only defines hotness in terms of
consecutive successes may be missing the point.
Larkey, Smith, and Kadane suggest a different definition of hotness. A bas-
ketball player who scores 10 points in 2 minutes will be thought of as being
“on fire” even if he or she has missed a couple of shots. The clustering of mul-
tiple successes into short periods of time is another way to run hot and cold,
even if the percentage of successes remains relatively constant. The Larkey
study of the NBA found several examples (notably Vinnie “The Microwave”
Johnson) of players who are streaky.
In the book Curve Ball, Albert and Bennett find evidence that the batting
averages of some batters fluctuate significantly over the course of a season.
This is not what most people mean by the hot hand, but it is another ex-
ample of the binomial model not always working. In Mathletics, Wayne Win-
ston notes that when the at bats are adjusted for park factors and pitching
matchups, the appearance of streakiness disappears.
Bocskocsky, Ezekowitz, and Stein collected data from the 2012-13 NBA
season and found evidence of several aspects of the hot hand. Using optical
tracking data, they found average success rates for shots based on player
ability, distance, angle, closeness of defenders, shot clock, and other variables.
They then monitored each player’s hotness by comparing recent success rates
to the average success rates. Thus, a player who had made two out of three 3-
168 Sports Math

pointers with a hand in the face would be hotter than someone who had made
four wide open layups in a row. By their measure of hotness, hot players are
more likely to (1) take the team’s next shot, (2) take a harder shot, (3) be more
closely guarded, and (4) make the next shot. This shows that (1) teammates
defer to the hot player, (2) the player is “feeling it” and gets overconfident,
(3) the defense adjusts to stop the hot player, and (4) the player is shooting
better than normal. Note that (2) cancels the effects of (4), which explains
the inability of researchers to find long sequences of successes.

Runs Tests
Other than taking my word for which of the three sequences in the “Hot
hands” section came from a coin flipping model, how can you tell? A statistical
test called the Wald-Wolfowitz Runs Test helps us evaluate sequences of
successes and failures. Define a run to be a sequence of the same letter. For
example, SSFSFFF has 4 runs: two Ss, one F, one S, three Fs. The Runs Test
uses the number of runs to analyze a sequence. Suppose a sequence (of at bats,
or field goal attempts, or wins/losses) of length n includes s successes and f
failures. For large n, the number of runs is approximately normal with mean
µ and standard deviation σ where

2sf (µ − 1)(µ − 2)
µ= + 1 and σ 2 =
n n−1
and the empirical rule is used to evaluate the likelihood that the sequence
comes from an independent and identically distributed process.
Each of the sequences above has length n = 30, s = 15 Hs, and f = 15
Ts. Then µ = 16 and σ 2 = 210 29 so that σ ≈ 2.7. Any number of runs outside
the interval µ ± 2σ = 16 ± 5.4 is suspicious. Thus, 10 or less and 22 or more
are suspect. The first sequence has 16 runs, exactly equal to the mean. The
second sequence has 21 runs, which is almost statistically significant. The
third sequence has 8 runs, which is three standard deviations below the mean
and therefore significantly low.
The Runs Test does not identify the second sequence as significantly dif-
ferent from independence. Looking at the lengths of the runs can accomplish
this. The first sequence has 16 runs. Half of them should be of length 1, half
of the remainder of length 2, and so on. We expect 8 runs of length 1, 4 runs
of length 2, 2 runs of length 3, 1 run of length 4, and 1 run of length greater
than 4. The first sequence has 8 runs of length 1, 4 runs of length 2, 2 runs
of length 3, and 2 of length 4. The second sequence has 14 runs of length
1, 3 runs of length 2, and 3 runs of length 3. This is not a good match. A
chi-square goodness of fit test quantifies the mismatch.
Another way to test a sequence is to create a large number of random
 Randomness in Sports 169

permutations and collect statistics. For example, create a random sequence of

15 Hs and 15 Ts. Do this a million times. How many of the sequences do not
have a streak at least four long? My simulation had 160,217. This is evidence
that the probability of a sequence having that characteristic of the second
sequence above (no runs of length 4 or more) is about 16%; i.e., not likely.

Joltin Joe and The Streak

In 1941, Joe DiMaggio got hits in 56 consecutive games. This broke the
old consecutive game streak of 44 games (it has since been revised to 45). The
large gap between best and second-best is one reason that DiMaggio’s record
is revered. Stephen Jay Gould wrote, “Thus Joe DiMaggio’s 56-game hit-
ting streak is both the greatest factual achievement in the history of baseball
and a principal icon of American mythology.” Equipped with some statistical
knowledge and play-by-play data from baseball-reference.com, let’s see how
impressive the record is.
During the streak, DiMaggio batted .408 (91 out of 223). If he batted 4
times in a game, the probability that he made 4 outs is .5924 = .123 so the
probability that he got a hit in a game is .877. He gets a hit in 56 straight
games with probability .87756 = .00064 or one time in 1600. Wow! That is
very unlikely. It is also a very bad analysis. What is wrong? The probability
of getting a hit in a game is wrong (see below), the calculation ignores the
fact that he had more than one chance to start the streak, and the calculation
implicitly assumes that his at bats were independent. Let’s do a little better.
The value of .877 for getting a hit in a game is too high. We used his
batting average, which ignores walks, so we assumed that he had four at bats
in every game without ever walking. During the streak, DiMaggio actually
had 246 plate appearances, so he got a hit in 37% of his plate appearances.
Using this figure in place of .408, his probability of getting a hit in a game
drops to 84%. (In fact, in 1941 he got a hit in 82% of his games. In 1940, he
got a hit in 86% of his games.)
A quick tangent: on May 15, 1941, both DiMaggio and Ted Williams
started the longest hit streak of their major league careers. Even though
Williams hit .406 for the entire 1941 season, his streak was only 23 games
long. Why did DiMaggio and not Williams have the longest hitting streak? In
the game that ended Williams’s streak, he walked three times. By contrast,
DiMaggio walked three times in the last thirty games of his streak in situa-
tions in which he had not yet achieved his hit. Williams, by the way, has the
major league record for most consecutive games reaching base, a staggering
84 straight games in 1949.
DiMaggio did not like to walk, and during the streak the pitchers were
under pressure to “play fair” and pitch to him. In addition, DiMaggio rarely
170 Sports Math

struck out. During the streak, he struck out only five times, with no strike-
outs in the last 32 games of the streak! This means that in almost every
plate appearance DiMaggio hit a fair ball. This makes DiMaggio an excellent
candidate for a long hitting streak.
Back to estimating the probability of DiMaggio’s streak happening: we
need to ask the question more precisely. That is, if you want the probability
of hitting in those exact 56 games, then .8456 ≈ .000057 is reasonable (and is
impressively small). However, if you want the probability that he would have
the streak at some point in 1941, multiply by 25 (the number of games in
which he did not get a hit in 1941, and therefore the number of chances he
had to start a new streak). We’re now up to 0.0014, but still well less than one
percent. If you want the probability that DiMaggio would have such a streak
at some point in his career, multiply by 5 or so; the probability is still less than
one percent. (Such probabilities will be explored in the exercises.) Numerous
estimates have been made for the probability that somebody sometime in
the history of baseball would have a 56-game hitting streak; two percent is a
common choice.
As Stephen Jay Gould said, the streak is a great achievement. However,
consider two more facts. After going hitless in game 57 (thanks to two great
defensive plays by the opposing third baseman, Ken Keltner), DiMaggio had
hits in the next 16 games to make it 72 out of 73! And, in 1933 DiMaggio set
a minor league record by hitting in 61 straight games!

Not Following the Rules

All of the calculations in the previous section assumed that DiMaggio’s
at bats were independent events. This cannot be true, but how much does
the assumption affect the calculations? Trent McCotter explored this ques-
tion in a paper with the provocative title of Hitting Streaks Don’t Obey Your
Rules. McCotter simulated 50 years of baseball using the actual averages and
schedules of the players.
The simulations gave interesting results. On the average, McCotter’s sim-
ulations produced (over the 50 simulated years) 49 players with streaks of at
least 25 games. In the real 50 baseball seasons, 62 players had streaks of 25
games or more. The table shows the simulation averages and actual numbers
for other lengths of streaks.
TABLE 8.3: Simulated vs Actual Streaks
length sim avg actual
25+ 49 62
30+ 10 19
35+ 2 5
 Randomness in Sports 171

Twice as many players as predicted by the independence model are achiev-

ing long hitting streaks. This doesn’t prove anything, but it does might make
you question calculations based on the independence assumption. Incidentally,
note how few players reach 35 games in a row, and recall that Joe DiMaggio’s
streak was 56 games!
Here is, to my mind, the most convincing evidence. In the 50 years, 4 real
baseball players had streaks of length exactly 29 games. McCotter’s simula-
tions produced 4 or more streaks of exactly 29 games about 25% of the time.
Nothing significant here. However, 9 real ballplayers had streaks of exactly
30 games. This many 30-game streaks never happened in McCotter’s 1000
simulations.
Real baseball players had longer streaks than their simulated (i.e., inde-
pendent) counterparts, but they were especially likely to extend a 29-game
hitting streak to 30. This looks like the result of human psychological (and
non-independent) effort. A related fact involves batting averages. From 1950-
2014, 89 players had a season batting average of .296, 104 players had an
average of .297, 82 players had an average of .298, 60 players had an average
of .299, and 178 players had a batting average of .300!

BABIP and DIPS

A baseball pitcher who gives up several hits in a row and is replaced by
another pitcher is said to have been “knocked out of the box.” Sometimes,
the phrase seems too harsh, as hits can be swinging bunts, ground balls that
sneak between fielders, and soft fly balls that barely clear the infield. Voros
McCracken found a way to quantify the bad luck that dogs some pitchers. He
found that approximately 30% of fair balls in play (not home runs) fall for
hits. Pitchers who give up hits on more than 30% of balls in play are likely
having a spell of bad luck that will disappear soon.
This leads to the BABIP (batting average on balls in play) statistic,
H − HR
computed as for a player with H hits in AB at bats,
AB − K − HR + SF + SH
HR home runs, K strikeouts, SF sacrifice flies, and SH sacrifice hits.
Research on BABIP indicates that the main control a pitcher has over his
BABIP against is whether batters tend to hit fly balls or ground balls against
him. Presumably, average velocity on batted balls (data that has become avail-
able recently) will also be found to affect BABIP against. Similarly, batters
have control over BABIP through percentage of ground balls, line drives, and
fly balls, as well as average velocity of batted ball.
Defense also plays an important role in BABIP, as a well-positioned or
outstanding individual defense will turn hits into outs. This insight leads to
DIPS, or Defense-Independent Pitching Statistics. McCracken’s version uses
172 Sports Math

strikeouts, walks, hit batsmen, and home runs allowed: the aspects of the game
that a pitcher can actually control. (Catchers do have a large effect on how
umpires call balls and strikes, so perhaps this also needs to be modified.) The
formula is a little messy because McCracken wanted it to be on the same scale
as earned runs allowed, but a simplified version is 3.2+(13HR+3W+3HBP−
2K)/IP.

Random Thoughts
In his book Mathletics, Wayne Winston tells of a statistical analysis gone
bad. Historically, when college basketball games have been fixed the players
are not asked to lose, but to win by less than the spread. Suppose the spread
is 15 points. Then the favorite could win by 10, make the gamblers happy, and
still win the game. Statistically, we might expect the results of honest games to
form a bell curve with mean 15 and standard deviation 10. About 34% of the
results would fall between 5 and 15 points (the empirical rule) and about 42%
of the results would fall between 1 and 14 points (inclusive). By symmetry,
about 42% of the results would fall between 16 and 29 points. However, a
study showed that college basketball games with large point spreads were not
symmetric, with 46.2% falling below the spread and 40.7% above the spread.
This result is consistent with games being fixed. However, it is also consis-
tent with the favorite not playing as hard, the favorite pulling the starters out
of the game earlier, and other non-sinister explanations. Investigators looked
at how the spread in these games changed over time, as a fixed game would
attract large bets on the underdog and would therefore cause the point spread
to decrease. In fact, it went the other direction, and that may be what caused
the asymmetry.
The lesson is to keep looking at the data, testing it in different ways to
learn as much as possible.
An important idea to keep in mind is sampling bias (or selection bias)
in which a bias is caused by circumstances. In looking at the play-by-play for
Joe DiMaggio’s hitting streak, we find that in situations in which he was 0-3,
he got hits in a remarkable 8 out of 9 at bats. Think about it for a second: he
usually had 4 at bats during the streak, and we are only looking at games in
the streak in which he always got a hit! The remarkable .889 batting average
is sampling bias caused by restricting ourselves to games in the hitting streak.
Sampling bias can be subtle. You should be wary any time that you are
limiting your data; you may unwittingly introduce a bias.
Here is one more interesting fact from Joe DiMaggio’s hitting streak. He
struck out only five times in those 56 games. In game 14, the strikeout occurred
in his last at bat of the game. In game 16, he followed the strikeout with a
 Randomness in Sports 173

double. In games 20, 23, and 24, he followed the strikeouts with home runs!
This does not look like independent at bats.

Exercises
In these exercises, T refers to thinking problems, conceptual problems requir-
ing no calculations. C refers to problems requiring significant calculations or
calculus. P refers to projects; these are ideas for further investigation (hints
and resources are at the book’s web site).

8.1 For each division in American League baseball in 2014, compute the mean
and standard deviations of wins. East: 96,84,83,77,71; Central: 90,89,85,73,70;
West: 98,88,87,70,67. Which division was the best? Which division had the most
balance?
8.2 If a basketball player’s points have mean 30 and standard deviation 4, in how
many games of an 82-game season would you expect the player to score (a) more
than 34 points; (b) less than 22 points; (c) more than 42 points?
8.3 Two running backs average 4 yards per carry. If player A has standard devi-
ation 2 yards per carry and player B has standard deviation 4 yards per carry,
which player is more likely to gain (a) more than 10 yards; (b) less than 0 yards;
(c) between 3 and 5 yards?
8.4 Draw a histogram and describe the shape for each. (a) Points scored for
the 2014 Dallas Cowboys (17,26,34,38,20,30,31,17,17,31,31,10,41,38,42,44); (b)
points scored for the 2014 New England Patriots (20,30,16,14,43,37,27,51,43, 42,
34,21,23,41,17,9). Compare.
8.5 Draw a histogram and describe the shape for each. (a) Wins in the NFL in
2014 (12,9,8,4,11,10,10,7,11,9,3,2,12,9,9,3,12,10,6,4,12,11,7,5,7,7,6,2,12,11,8,6); (b)
wins in the NBA in 2014 (49,40,38,18,17,53,50,41,38,32,60,46,37,33,25,51,45,38,
30,16,67,56,39,29,21,56,55,55,50,45). Compare.
8.6 In college football games on a certain date, teams favored by 1, 2, 3, 4, and
5 points won by 12, lost by 4, won by 2, won by 12, and lost by 6, respectively.
How often did the favorite win against the spread? Find the differences of spread
minus result (e.g., the first difference is 1 − 12 = −11), and find the mean and
standard deviation of the differences.
8.7 Given a standard deviation of 10 points, if a college basketball team is favored
to win by 10 points, what percentage of games should it win?
8.8 A team wins 60% of its games. Assuming independence, (a) compute the
probability of the team losing 3 out of 4; (b) use the empirical rule to estimate
the probability that the team wins 55 or less out of 100.
8.9 A team wins 80% of its games. Assuming independence, (a) compute the
probability of the team losing 3 out of 4; (b) use the empirical rule to estimate
the probability that the team wins 76 or less out of 100.
174 Sports Math

8.10 A team wins 35% of its games. (a) Find the probability that it loses its first
21 games. (The 1988 Orioles did this.) (b) Find the probability that this team
would lose its first 21 games at least once in 400 tries. (c) How does this compare
to estimates of the likelihood of a 56-game hitting streak?
8.11 On June 14, 2015, Andre Iguodala of Golden State (the eventual MVP of the
Finals) made 2 out of 11 free throws in game five of the NBA Finals. (a) Using
his 70% lifetime free throw percentage, find the probability that he makes 2 or
fewer free throws out of 11. (b) Repeat using his season free throw percentage
of 60%. (c) How unlikely was this performance?
8.12 A baseball teams wins 56% of its games. (a) Find the expected number of
wins in a 162-game season. If it takes 91 wins to make the playoffs and 97 wins
to finish in first, use the empirical rule to estimate the probability of (b) making
the playoffs; (c) finishing in first.
8.13 In the discussion after Example 8.3, one method of combining win per-
p1 (1−p2 )
centages p1 and p2 into a win probability was given: e1 = p1 (1−p 2 )+p2 (1−p1 )
.A
simpler method is e2 = 0.5 + p1 − p2 , and a third method (called the “James
p1 −p1 p2
log-5” method) is e3 = p1 +p 2 −2p1 p2
. Show that e3 = e1 . Compare e1 and e2 for
the following values. (a) p1 = .55, p2 = .52; (b) p1 = .60, p2 = .57; (c) p1 = .80,
p2 = .77; (d) p1 = .5, p2 = .4; (e) p1 = .55, p2 = .45.
8.14 Sketch the Lorenz curve, find the Gini index, and compute the entropy for
5-person basketball teams with the following scoring breakdowns. (a) {40, 22, 8,
6, 4}; (b) {20,19,18,17,16}
8.15 Sketch the Lorenz curve, find the Gini index, and compute the entropy for
small leagues with the following win breakdowns. (a) {18, 16, 14, 12, 10, 8, 6, 4,
2, 0}; (b) {11, 10, 9, 9, 9, 9, 9, 8, 8, 8}
8.16 For a team with five players and using the computational technique in Ex-
ample 8.4, find the maximum and minimum Gini index.
8.17 If the luck variance is 40, the skill variance is 400, and the actual variance
is 100, find the fraction p of skill.
8.18 (a) If batting averages have mean .270 and standard deviation .030, estimate
the probability of a player batting .300. (b) Repeat with a standard deviation of
.015. (c) Explain why the second probability is lower.
8.19 Find the means and standard deviations of the top 10 runners in the
Olympics women’s marathons from 1984 to 2012. Are the winning times im-
proving? Are the mean times improving? Are the standard deviations decreas-
ing? Discuss whether the runners are improving.
8.20 Kobe Bryant made 45% of his shots in his career. In 2011-12 in clutch
situations (last 30 seconds, less than 3-point score differential) he made 5 out of
22. Find the probability that with p = .45 there would be 5 or fewer successes in
22 tries. Given that everybody in the arena knows that Kobe will take the last
shot, is it fair to use his lifetime field goal percentage?
8.21 Use the runs test to test each of the following for independence. (Each has 10
S’s and 10 F’s.) (a) SFSSSSFFSFSFFSSFFFSF; (b) SSSFFFFSSFSSSSFFFFFS
8.22 Use the runs test to test each of the following for independence. (The
numbers of S’s and F’s are not equal.) (a) SSSSFSSSFSSSSSFSSSSS; (b)
SFFSFFSFFFSFFFSFFFFS
8.23 In the 2015 NBA Three-Point Contest, Steph Curry won, recording the
 Randomness in Sports 175

following sequence of makes (Y) and misses (N). Use the runs test to test for sig-
nificance. If he is really a 50% three-point shooter, compute the probability that
he would make at least 20 out of 25. NNYYYYYYYNNYYYYYYYYYYYYYN
8.24 In the 2015 baseball Home Run Derby data, batters hit 159 homers in
451 pitches. The sequence of homers formed 206 runs (streaks). Compute the
expected number of runs and test for significance. The expected percentages of
homer droughts of length 1, 2, 3, and so on are 35, 23, 15, 10, 6, 4, 3, 2, 1, 1.
Compare to the actual values of 34, 25, 12, 14, 5, 4, 2, 1, 2, 1. The expected
percentages of homer streaks of length 1, 2, 3, and so on are 65, 23, 7, 3, 1, 1.
Compare to the actual values of 68, 22, 4, 3, 0, 3. Is there evidence that the
hitters were unusually streaky? Winner Joc Pederson hit six homers in a row.
What is the probability of him doing that? Why is this small probability not
necessarily evidence of hot hands?
8.25 If a player gets hits 35% of the time, find the probability of getting at least
one hit in (a) 4 at bats; (b) 5 at bats.
8.26 If a player gets hits 50% of the time, find the probability of getting at least
one hit in (a) a game of 4 at bats; (b) 56 straight games with 4 at bats each.
8.27 In Joe DiMaggio’s streak, games shortened by rain sometimes limited his at
bats. With a batting average of .370, compute the probability of getting at least
one hit in five consecutive games with 20 at bats distributed in the following
ways. (a) 4, 4, 4, 4, 4; (b) 5, 2, 5, 5, 3
8.28 In the text’s analysis of DiMaggio’s streak, the 56-game probability of .8456
is multiplied by 25 to account for the multiple opportunities for DiMaggio to
start a streak. The true probability is 1 − (1 − (.84)56 )25 . Explain why this is the
correct value, compute it, and compare to the value in the text.
8.29 To see why the issue raised in exercise 8.28 is important, consider taking a
basic probability of 0.001 for some event, and asking for the probability that it
occurs at least once in 2000 tries. What happens if you just multiply by 2000?
Explain why this cannot be correct, and then compute the exact probability.
8.30 Compute BABIP against (assume no sacrifice hits) and DIPS for the
given data for Clayton Kershaw and Greg Maddux. How do the BABIP-
against values compare to the average BABIP of .300? Is there any ev-
idence that either pitcher was luckier in one season than the other?
IP AB H HR BB HBP SO
Kershaw 2014 236 851 164 11 54 3 232
Kershaw 2013 237 828 170 16 68 5 229
Maddux 1994 202 734 150 4 34 6 156
Maddux 1993 267 999 228 14 59 6 197
8.31 Compute BABIP for the given data for Tony Gwynn and Jim Thome.
How do the BABIPs compare to the average BABIP of .300? Is there
any evidence that either hitter was luckier in one season than the other?
AB H HR SO SF SH
Gwynn 1994 419 165 12 19 5 1
Gwynn 1995 535 197 9 15 6 0
Thome 2001 526 153 49 185 3 0
Thome 2002 480 146 52 139 6 0
8.32 T For Bernoulli trials with p ≥ .5, fill in the blank with “larger” or
“smaller” and explain: the larger p is, the —– the standard deviation is.
176 Sports Math

8.33 T The paradox of skill in Figure 8.10 assumes a fixed limit of human
performance, so that skill levels compress. Given the fact that track, swimming,
and other records of measurable performance still improve, discuss whether the
fixed limit of performance is a reasonable assumption.
8.34 T Discuss the ideal balance of skill and luck in a sport.
8.35 T In recent years, the standard deviations of batting averages have not de-
creased. Two explanations could be: (a) baseball players are not close to the limit
of human performance; (b) batting average is not selected for, in that managers
care more about other skills. Discuss the relative merits of these explanations.
8.36 T Discuss the relative values of the great batting averages of Ted Williams
in 1941 and Wade Boggs in 1985.
8.37 T Explain why, assuming games are independent, the 2-3-2 and 2-2-1-1-1
schedules produce the same winners with the same probabilities. Explain why a
4-3 or 2-1-2-2 schedule would not be equivalent.
8.38 T According to Analyzing Wimbledon, after breaking serve men tennis
players hold serve 65.8% of the time, whereas after failing to break they hold serve
64.1% of the time. This looks like a violation of independence, but explain why
it could be an example of sampling bias. (Hint: different players have different
skill levels.)
8.39 T Explain why the probability of each sequence of Hs and Ts in the “Hot
Hands” section has the same probability of occurring in coin flipping. Given this,
what is the runs test looking for?
8.40 T Describe the feeling of being “in the zone.” Discuss whether this feeling
is the result of a string of successes, or whether the string of successes is the
result of the feeling. In other words, which causes which?
8.41 T For each of the following, explain what is suspicious about each se-
quence as the product of fair coin tossing. (a) HHHHHTHHHHHHTTHHHHH;
(b) HTHTHTHHTHTTHTHHTHTTH; (c) HHTTTHHHTTHHTTTHHTT
8.42 T In the debate about whether Hot Hands exists, describe what is to you
the most important piece of evidence.
8.43 T While nobody has approached DiMaggio’s record of 56 straight games
with a hit, Derek Jeter once got hits in 59 out of 61 games, and Johnny Damon
got hits in 60 out of 63 games. Do these “near-misses” make DiMaggio’s record
seem more or less heroic? Discuss.
8.44 T Discuss whether the difference between the numbers of hitting streaks
of lengths 29 and 30 is convincing evidence of independence being violated.
8.45 T Explain why batters have more control over BABIP than do pitchers
over BABIP against.
8.46 T In 1981, the Clemson University football team went 12-0 and won the
national championship. They wore special all-orange uniforms in three impor-
tant games that year. Discuss the validity of the statement that Clemson was
unbeatable in its all-orange uniforms.
8.47 C Draw a histogram and describe the shape for each. (a) Points scored
and (b) points allowed in games for the 2014-15 San Antonio Spurs; (c) runs
scored and (d) runs allowed in games for the 2014 San Francisco Giants.
 Randomness in Sports 177

8.48 C Draw a histogram and describe the shape for each. (a) Home runs for
American League players in 2014 (players with at least 2); (b) Rebounds per
game for NBA players in 2014 (players with at least 2 rebounds per game).
8.49 C The pdf for a normally distributed random variable with mean µ and
2 2
variance σ 2 is f (x) = σ√12π e−(x−µ) /2σ . Take µ = 0 and σ = 1 and estimate
R1
−1
f (x)dx. Explain how this relates to the empirical rule.
8.50 C The shape seen in Figure 8.4 can be described by power law functions
of the form p = cx−k and exponential functions of the form f = ce−kx . Show
that ln(p) = ln(c) − kln(x) and ln(f ) = ln(c) − kx. Thus, if you have data (x, y)
that is from a power law function, a “log-log” plot of ln(y) versus ln(x) will be
a straight line. By contrast, if you have data (x, y) that is from an exponential
function, a “semi-log” plot of ln(y) versus x will be a straight line. Use log-log
and semi-log plots to determine which is power law and which is exponential
and determine the underlying equations. (a) {(1,20),(3,2.2),(5,0.8),(7,0.41)}; (b)
{(0,20),(2,2.7),(4,0.37),(6,0.05)}
8.51 C With a standard deviation of 10 points, how many points better than
the spread would you have to be to win 65% of your bets?
8.52 C Compute the probabilities in exercises 8(b) and 9(b) exactly, and com-
pare your answers to the values obtained from the empirical rule.
8.53 C A team wins 12 out of 20 games. Assuming that games are Bernoulli
trials with probability p, write an expression for the probability of winning 12
and losing 8. Think of this as a function of p, compute its derivative and find the
value of p that maximizes the probability. This is called a “maximum likelihood
estimate” of p.
8.54 C Find the Gini index for f (x) = x2 /100.
8.55 C Show that the maximum entropy for two players occurs when both
proportions are 0.5. Assuming that the maximum entropy occurs with equal
probabilities, show that the maximum entropy for n players equals ln(n).
8.56 C Find the fractions p of skill for the NBA in 2009-10, 1999-00, 1989-90,
and 1979-80 and comment on any trends.
8.57 C Find the expected length of the NBA Finals using the 2-3-2 schedule
versus the 2-2-1-1-1 schedule assuming (a) all games are independent with the
home team winning 65%; (b) all games are independent with the better team
(the first one playing at home) winning 75% of its home games and the lesser
team winning 55% of its home games.
8.58 C In the case of .200-hitter Bob Uecker batting .300 against Steve Carlton,
how many at bats would be needed for this result to be statistically significant?
8.59 C Pete Rose’s lifetime batting average was .303. Assuming independence,
calculate the probability that Rose would get (a) at least 13 hits in 30 at bats
against Greg Minton; (b) at least 16 hits in 36 at bats against Bob Friend.
8.60 C Find the probability that the better team wins a best-of-5 series assum-
ing that: the better team starts at home and wins 75% of its home games, the
lesser team wins 55% of its home games, winning the most recent game adds ten
percentage points to the chance of winning (e.g., from 75% to 85%), and (a) a
2-2-1 schedule; (b) 1-1-1-1-1 schedule.
178 Sports Math

8.61 C The 1965 New York Giants made a remarkable 4 out of 26 field goals in
1965. Compute the probability of a performance this bad or worse if the Giants
had an average kicker (the league made 334 out of 617 field goals in 1965).
8.62 P For NFL teams in 1985-2004, compute the Pythagorean wins for each
team with exponent 2.37 and the Massey win ratings and identify the best teams.
Determine how many times the best team won the Super Bowl. Has this pattern
continued since 2004?
8.63 P Repeat the Bill James simulation (described following Example 8.3)
using the current baseball divisions and schedule. How often does the best team
win?
8.64 P Find the Gini indices and entropy for all teams from the league of your
choice (NFL, NBA, ...). How do these measures of parity correlate to success?
8.65 P Simulate numerous seasons of the Luck-NBA to find the variance of
wins. Compare your answer to the binomial-based given used in the text.
8.66 P Show that there have been more 7-game World Series than a binomial
model would predict. Create a model that explains this phenomenon.

Further Reading
Analytic Methods in Sports by Severini gives a thorough development of
statistical methods used in sports analysis, with numerous interesting exam-
ples.
Nate Silver’s The Signal and the Noise explores various aspects of sepa-
rating real patterns from random occurrences in sports, politics, and life.
In “The Relationship Between Concentration of Scoring and Offensive Ef-
ficiency in the NBA,” Ruiz, Martinez, Lopez-Hernandez, and Castellano use
the Gini index to investigate the question of whether star-driven NBA teams
are more successful than balanced-scoring teams.
Hot Hand by Reifman covers many of the topics included here. Papers
by Tversky and Gilovich and Larkey, Smith, and Kadane are in Anthology of
Statistics in Sports edited by Albert, Bennett, and Cochran.
56 by Kennedy and Streak by Seidel give excellent accounts of Joe DiMag-
gio’s hitting streak.
Bill James’ essay “Underestimating the Fog” is a helpful counterpoint to
some of the no-hot hands claims.
Triumph and Tragedy in Mudville collects many of Stephen Jay Gould’s
essays on baseball.
The Information by Gleick introduces many of the ideas in information
theory, including entropy. See also Fortune’s Formula by William Poundstone.
Chapter 9
Sports Strategies

Introduction
The opening session at the
2014 Sloan Sports Analytics
Conference in Boston featured
Bill James (hero of Chapter
7), Hall of Fame NBA coach
George Karl, Houston Rockets
General Manager Daryl Morey,
best-selling author Nate Silver, FIGURE 9.1: Kevin Kelley?
and Kevin Kelley. In case the
name Kevin Kelley does not ring a bell, he is a high school football coach
in Arkansas. His teams have won multiple state championships. Still puzzled
by his star billing? Kevin Kelley is the Billy Beane of football.
Kelley took a serious look at some of the numbers we will develop in this
chapter and dared to ignore convention and implement innovative strategies.
His teams, essentially, never punt. They do not try field goals. After touch-
downs, they go for two points ... and then try an onside kick. In one important
and otherwise close game, Kelley’s team led 28-0 before the other team ran an
offensive play. This is why Kevin Kelley is a star! Instead of saying that the
numbers are nice but nobody is crazy enough to play this way, Kelley puts it
on the field and sees what works.
Mathematical analysis sometimes verifies the efficiency of standard sports
practices. The challenge for analysts occurs when the mathematics suggests a
better way of doing business. When a Billy Beane or Kevin Kelley is willing
to put the new ideas into practice, the believers and the skeptics learn what
works and what does not. Successes are copied, and failures are re-analyzed.
The numerical logic behind the strategies in sports is the topic of this
chapter. We will attempt to answer the following questions. Is Kevin Kelley
crazy? What about Bill Belichick? Is bunting a good strategy in baseball?
What is the best strategy for penalty kicks? How much does a team’s batting
order matter? What do “leverage” and “game control” mean? What does
John Nash have to do with sports? How important is it for teams to have star
players? At what age do baseball players hit the most home runs?

179
180 Sports Math

Don’t Punt, John!

Your team faces fourth down with two yards to go at midfield. Which
option do you choose, go for it or punt? You already know Kevin Kelley’s
answer, but let’s see in what way it is correct.
We start with some numbers from pro-football-reference.com. In 2014,
NFL teams went for it on fourth down and 2 or 3 on 83 occasions, and were
successful 53% of the time. The average gain was 6 yards. So, let’s say your
team is successful 53% of the time, and you either get a first down at your
opponent’s 40 or they take over the ball at the 50. The alternative is to punt.
In 2014, the 107 punts from the 50 averaged 40 yards with a return of 6 (no
information on how many touchbacks). Let’s say the outcome of a punt is for
the opponent to take over on its 16.
How do we evaluate which is the better strategy? One way is to compute
an expected value of the score. The relevant research, amazingly enough, dates
back to Virgil Carter, a quarterback for the Cincinnati Bengals who published
a mathematics paper in the journal Operations Research in 1971. The idea is
to track every drive for several seasons and note the next score. For example,
a team starts a drive on its own 20 and scores 7 points; another team starts
a drive on its own 20, does not score, but its opponent scores 3 points on its
next possession. At this point, 2 drives from the 20 have netted +4 points
(7-3) for an average of +2 points. Doing this for every yard line and several
years gives a good idea of how many points a position on the field is worth.
It turns out that the data is very close to forming a straight line connecting 6
points when 1 yard from the goal line (ask the 2015 Super Bowl teams why it
is not 7) to -1 point (the other team is more likely to score next) at 100 yards
from the goal line. The function p(x) = 6 − .07x gives an approximation of
the expected points when x yards from the goal line.

Example 9.1 Compute the expected values for going for it and punting from
the 50 on 4th down and 2.
Solution Recall that expected value equals the sum of probability times value
for all possible outcomes. Going for it, we have two possible outcomes. Making
a first down has probability 53% and value p(40) = 3.2, the expected num-
ber of points from a first down at the opponent’s 40. Getting stopped has a
probability of 47% and value −p(50) = −2.5, the negative of the opponent’s
expected points from a first down at the 50. After a punt, the expected value
is −p(84) = −.12. The expected values are
Going For It: .53 ∗ 3.2 − .47 ∗ 2.5 = 0.521
Punting: −0.12
The expected value is larger (and positive) for the strategy of going for it. (As
Kevin Kelley could have told us.)
 Sports Strategies 181

A coach would be considered a reckless gambler going for it at the 50.

What about going for it inside your own territory?

Example 9.2 Compute the expected values for going for it and punting from
your own 30 on 4th down and 2.
Solution We will use the 53% success rate from Example 9.1, and assume
that the options are first down at the 40 or the opponent’s ball at your 30.
From pro-football-reference.com, punts from the 30 in 2014 averaged 43 yards
with a 9-yard return, giving the opponent the ball at its 36.
Going For It: .53 ∗ p(60) − .47 ∗ p(30) = −0.88
Punting: −p(64) = −1.52
The expected value for going for it is now negative, but it is less negative than
the punting value. Therefore, you should go for it!

This is starting to look crazy. Let’s phrase that differently: either pro
and college football coaches are choosing ineffective strategies, or there is
something wrong with these calculations. The importance of Kevin Kelley is
that he has shown us that, at his high school level, the calculations are largely
correct. It is important to note that the lower quality of punting in high school
makes Kelley’s refusal to punt even sounder.
Of course, the calculations grossly oversimplify the decisions that coaches
make. Offensive and defensive match-ups vary, the score and time remaining
are important considerations, and so on. If teams start going for it routinely,
the success rates will change. Nevertheless, Examples 9.1 and 9.2 raise inter-
esting questions.
In a side note, Virgil Carter had a starring role in a strategy situation
that changed professional football (as told in the book Newton’s Football ).
Carter was backup to Greg Cook, who was having a sensational rookie season
running Bill Walsh’s offense in Cincinnati. When Cook was injured, Carter
took over but did not have the arm strength to make the necessary throws. In
desperation, Walsh shifted the offense to a Carter-friendly short passing game
that was the genesis of the West Coast offense that Walsh perfected in San
Francisco.

Bill Belichick’s Gambles

If you are a long-time New England Patriots or Indianapolis Colts fan, you
recognized the scenario in Example 9.2. In 2009, with a 34-28 lead and only
two minutes remaining, Patriots Coach Bill Belichick went for it on fourth
down and two to go. The Patriots did not get the first down and lost the
game 35-34 on a late touchdown pass. Example 9.2 indicates that the decision
182 Sports Math

was not dumb, especially given the Patriots’ 75% rate of converting fourth
downs.
In the 2012 Super Bowl, Belichick’s Patriots led 17-15, but the New York
Giants had the ball at the New England 6 and the clock was approaching
one minute remaining. Belichick ordered his defense to let the Giants score a
touchdown, gambling that his offense could score a touchdown in one minute
to retake the lead. They did not, and the Giants won. The key issue here is
not expected points, but the probability of winning the game. A field goal
would have won the game for the Giants and, other than a highly unlikely
fumble or missed field goal, there was no scenario in which the Patriots could
win the game if the Giants had the ball.
As with expected points, win probabilities can be computed if enough data
is available. Unlike expected points, win probabilities require several variables.
With expected points, it makes some sense to say that teams average 2.5 points
when starting 50 yards from the goal line. The score and amount of time left
are not necessarily important (an obvious exception being if only two seconds
are left). However, it makes no sense to ask what the probability of winning is
when you start 50 yards away; you have to know the current score and time
remaining, also. This makes win probabilities harder to approximate; there
is no simple p(x) formula. As Advanced NFL Stats (later renamed Advanced
Football Analytics) computed it, the Giants had a 94 percent chance of win-
ning before the play. After the Patriots let them score, the probability of the
Giants winning dropped to 85 percent. The Patriots still had little chance to
win, but the odds were better.
Interesting plays seem to follow Belichick, because in the 2015 Super Bowl
there was a near replay of the 2012 game. The Patriots led the Seattle Sea-
hawks 28-24 with just over one minute remaining, but Seattle had the ball
at the New England 1. Instead of letting Seattle score or calling a timeout
to preserve time, Belichick allowed the clock to run. In a stunning finish, the
Patriots won the game when Malcolm Butler intercepted a pass on the next
play. In this case, Seattle had an 88 percent chance of winning before the play.
This would have dropped to 77 percent if the Patriots had called timeout and
Seattle scored a touchdown. (It would have been less than the Giants’ 85 per-
cent because in this case Seattle would have led by three points, so that a
Patriots field goal could have tied the game.)
The evidence is that Belichick made the right call in the first two cases,
and lost, and made the wrong call in the last case, but won. Probabilities are,
after all, only probabilities.
What about Seattle’s decision to try a pass? Here are some numbers from
pro-football-reference.com. On plays with two or less yards to go for a touch-
down, in 2014 teams ran the ball 333 times and threw 208 times. Runs pro-
duced touchdowns 54.1% of the time, and turnovers 1.5% of the time. Passes
produced touchdowns 52.4% of the time, and only one turnover (a fumble).
Thus, calling for a pass was neither unusual nor dumb, especially given the
amount of time remaining. The outcome was unfortunate for Seattle.
 Sports Strategies 183

The Value of a Play

The concepts of expected points and win probability give us different ways
of evaluating plays and players. For example, a team completes a 70-yard pass
from its own 20 to its opponent’s 10. Instead of calling this 70 yards, we can say
how many points it is worth. Before the play, the offense had expected points
of p(80) = 0.4 and after the play expected points increase to p(10) = 5.3. The
play is worth 5.3 − 0.4 = 4.9 points.
Three plays later, the team’s running back scores a touchdown from the
1-yard-line. He gets credit for the touchdown, but did not actually add much
to the expected score. Using the basic expected points formula, he added 1
point (from an expected 6 to an actual 7).
Win probabilities could tell a different story. If the 70-yard pass play was
at the end of a 35-0 blowout, then the win probability would not change; the
probability is 100% at 35-0, and doesn’t get any higher at 42-0. If it were at the
end of a game that was tied, it would increase the probability of winning from
about 50% to about 90%. The WPA (win probability added) points would be
40, reflecting a play with a large influence on the outcome of a game.
Notice that the play would receive a large number of WPA points even if
the team lost the game due to a subsequent fumble or missed field goal.
If the expected number of points follow a straight line, a counterintuitive
conclusion follows, as shown in Carroll and Palmer’s book The Hidden Game
of Football.

Example 9.3 Use the expected points formula p(x) = 6 − .07x to compute
the cost of losing a fumble at (a) your own 10-yard-line; (b) your opponent’s
10-yard-line; (c) x yards from the goal line.
Solution (a) At x = 90 yards from the goal line, expected points equal
p(90) = −.3. The opponent gets the ball 10 yards away with expected points
p(10) = 5.3 which is an expected −5.3 points from your perspective. The
difference is −5.3−−.3 = −5.0 points. (b) At x = 10 yards away, you drop from
an expected 5.3 points to an expected .3 points, a difference of .3 − 5.3 = −5.0
again. (c) For any x, you go from p(x) to −p(100 − x), with a difference of
−p(100−x)−p(x) = −[6−.07(100−x)]−[6−.07x] = −6+7−.07x−6+.07x =
−5, exactly as we computed in parts (a) and (b).

Is it true that a fumble costs 5 points no matter where on the field it occurs?
It certainly doesn’t feel that way to fans. Perhaps the emotional impact of a
fumble near the goal line changes the expected points formula. Otherwise, the
formula forces us to conclude that a fumble is a fumble. As you will show
in the exercises, any linear function for p(x) will force the same qualitative
conclusion.
184 Sports Math

Markov Chain Models

In this section, we introduce a mathematical technique that proves valu-
able for analyzing some sports. Baseball is an orderly sport that lends itself
to mathematical analysis. There are a small number of states, or situations,
that describe all possibilities in an inning. There are eight base situations (no
runners, runner on first, runner on second, runners on first and second, and so
on) which can be paired with the 3 out possibilities to make 24 states. Adding
the “inning over” possibility makes 25 states. An example of a state is “1 out,
runners on second and third.”
A Markov chain is a set of states and transition probabilities that de-
scribe movement between states. There is, as usual, an assumption of indepen-
dence, that past results do not affect the probabilities. By studying Markov
chains in general, mathematicians have discovered important formulas to sim-
plify calculations. Here, we will work with a small example to see how this
might work.
We invent a simple version of cricket/baseball, let’s call it OneBase, in
which each batter either gets a single, a homer, or an out. A single puts the
batter on base; if there was a runner on base, he scores. A homer counts as one
run (two if a runner is on base). An out is like a strikeout; a runner does not
advance. An inning consists of two outs, and there is only one base other than
home. There are five states: (1) 0 outs, 0 on base; (2) 0 outs, 1 on base; (3) 1
out, 0 on base; (4) 1 out, 1 on base; (5) 2 outs, inning over. This last state is
called an absorbing state since once the second out is recorded nothing else
happens; the process has been absorbed into the fifth state forever.
Assume that a homer occurs with probability .1, a hit with probability .2,
and an out with probability .7. The first task is to compute the transition
probabilities, the probabilities of changing from one state to another. Start
with T11 , the probability of starting in state 1 (0 outs, 0 on base) and staying
in state 1. The only way this can happen is if the batter hits a homer, which
happens with probability .1. So T11 = .1. Next, T12 is the probability of
starting in state 1 and ending in state 2; this happens if the batter hits a
single, with probability T12 = .2. The transition from state 1 to state 3 occurs
if the batter makes an out, so T13 = .7. It is not possible to move directly (in
one batter) from state 1 to state 4 or state 5, so T14 = T15 = 0.
Even with our reduced set of states, this can get tedious. We store all of
our probabilities in a matrix. The five numbers we just computed go in the
first row of the matrix. So Tij will refer to the number in the i-th row and
j-th column. Note that this means that if you add the numbers in any row,
you will get a sum of 1. We have
 Sports Strategies 185

From, To 0 out, 0 on 0 out, 1 on 1 out, 0 on 1 out, 1 on over

0 out, 0 on .1 .2 .7 0 0
0 out, 1 on .1 .2 0 .7 0
1 out, 0 on 0 0 .1 .2 .7
1 out, 1 on 0 0 .1 .2 .7
over 0 0 0 0 1

which without the labels can be summarized in the matrix T given by

 
.1 .2 .7 0 0
.1 .2 0 .7 0 
 
 0 0 .1 .2 .7
T = 
 0 0 .1 .2 .7
0 0 0 0 1

which is called the transition matrix. Take a second to follow the logic in
each line of the matrix. The state changes with each batter, either through a
homer, a single, or an out. Thus, you see the three associated probabilities in
each line except the last. For example, the fourth row describes what happens
starting with 1 out and 1 on. A homer turns the state into 1 out and 0 on
(two runs score), a single leaves the state at 1 out and 1 on (one run scores),
and an out is the second and last out of the inning. The last row simply says
that when the inning is finished, it remains finished.
One quantity of interest here is runs scored. To see how to get runs in-
volved, consider the second row in which the initial state is 0 out and 1 on.
A homer (probability .1) scores 2 runs, and a single (probability .2) scores 1
run. The expected number of runs for one batter from this state is 2x.1 + 1x.2
= .4. The only way to score from states 1 and 3 is a homer, so the expected
number of runs for one batter from states 1 and 3 is .1. The expected runs
scored for one batter can be summarized in the vector

R = [r1 , r2 , r3 , r4 ] = [.1, .4, .1, .4, 0]

This was not hard to compute, but the quantity that we really want is the
expected number of runs scored from a given state for the entire inning. This
calculation will look like the matrix calculations in Chapter 5 for the Massey
ratings.
Start by assigning names to the expected number of runs from each state:
we can use a, b, c, d, and e. Think through the process of scoring runs from
the first state, 0 out and 0 on. Runs can be scored from the first batter (what
we computed above) or by scoring runs starting with the second batter. Then
186 Sports Math

a, the total expected runs scored, equals the initial value of .1 plus subsequent
runs scored from any of the other states. From state 1, the transition to state 1
has probability T11 , and from state 1 we expect to score a runs; the transition
to state 2 has probability T12 , and from state 2 we expect to score b runs; the
transition to state 3 has probability T13 , and from state 3 we expect to score c
runs; the transition to state 4 has probability T14 , and from state 4 we expect
to score d runs. All in all, we have

a = r1 + T11 a + T12 b + T13 c + T14 d

b = r2 + T21 a + T22 b + T23 c + T24 d
c = r3 + T31 a + T32 b + T33 c + T34 d
d = r4 + T41 a + T42 b + T43 c + T44 d

where the equations for b, c, and d are derived in the same way.
From here, you can get a computer or calculator to solve the equations for
you. (In linear algebra terms, the solution is (I − S)−1 R where I is the 4x4
identity matrix and S is the upper 4x4 submatrix of T .) The solution is
a = .52, b = 1.03, c = .23, d = .53.
The best scoring state is the second state with 0 out and 1 on, and the worst
scoring state is the third state with 1 out and 0 on. This is logical.
The same process can be done for baseball. There are various issues for
getting the best model possible. Instead of adding one state for “inning over”
you can allow runs to score on the last out by making four extra states cor-
responding to the last out made with 0 runs scoring, the last out made with
1 run scoring, and so on. Further tweaks can account for stolen bases, balks,
and other events that do not result in a change of batter. The probabilities in
the transition matrix can be estimated simply (find the proportion of at bats
resulting in a double and assume that all runners advance two bases) or more
fully (e.g., the probability of a double could depend on the current state).
Several of the statistics that can be computed from Markov chain models of
baseball agree closely with league statistics. This serves to validate the model,
which might be dubious given an assumption of independent at bats. With
the model in place, experiments can be run to test strategies. One interesting
result that has been explored by several researchers is the effect of batting
order in baseball. The bottom line is that it does not seem to matter much,
with the difference between best order and worst order being 30 to 40 runs
(3 or 4 wins). The general principle is that you want the best batters to bat
near the top of the order, so that they get the most at bats. As Bill James has
said, the main difference in most reasonable batting orders is in which innings
the runs will be scored.
 Sports Strategies 187

The Expected Runs Matrix

Once you have the average number of runs scored from each state, you
can analyze in-game strategies. From an expected runs matrix found online,
you can answer questions like the following, where the data comes from Tom
Tango’s website.

Example 9.4 Find the expected run values for a runner on first with 0 outs
and a runner on second with 1 out. Discuss whether a sacrifice bunt is a good
play.
Solution At tangotiger.net, the expected runs scored with 0 outs and a runner
on first from 1993-2010 is 0.941, while the expected runs scored with 1 out
and a runner on second is 0.721. The team is better off in the first situation,
so a sacrifice bunt that puts the team in the second situation is an ineffective
play.

There are numerous problems with the analysis in Example 9.4. The tables
give an average/expected number of runs over multiple seasons. If the batter
is below average (for example, a pitcher), the bunt may be a good play. This
is explored in the exercises. A different objection is that the goal may not be
to score the most runs. If it is a 3-3 game in the bottom of the ninth, the
team only needs one run to win. Instead of looking at the expected number
of runs, the probability of scoring one run is more relevant. Tango’s website
includes that information as well. Teams score from a runner on first, 0 out
state 44.1% of the time, compared to 41.8% of the time from the runner on
second, 1 out state. The bunt still looks like an inferior play, in general.

Win Probability and Leverage

The concept of win probability is straightforward and has been referred to
several times. Instead of looking at how many points or runs are scored from a
particular state, we look at the probability of winning from a particular state.
As previously discussed, this means we need to account for score and time
remaining. In baseball terms, this means looking at score and inning.

Example 9.5 Find the win probability for a runner on first with 0 outs and
a runner on second with 1 out, in the bottom of the ninth with the score tied.
Discuss whether a sacrifice bunt is a good play.
Solution At tangotiger.net, the win expectancy (probability) for innings 7-9
and run differentials of 0 or 1 are given. With 0 outs and a runner on first
188 Sports Math

in the bottom of the ninth with the score tied, the probability of the home
team winning is about 0.715. The win probability with 1 out and a runner on
second is 0.703. It is close, but the team is better off in the first situation, so a
sacrifice bunt that puts the team in the second situation is an ineffective play.

Win probabilities allow us to measure how critical a play is, or how much
leverage a play has on the outcome. A simple form of the Leverage Index
can be computed using the situation in Example 9.5 with the batter swing-
ing. There are multiple possible outcomes. Let’s say that this pitcher-batter
matchup produces home runs 10% of the time, singles 20% of the time, and
simple outs 70% of the time (no double plays, or runners advancing). The
team starts with a win probability of .715, which is raised to 1 (they win!) on
a home run. This is a change in win probability of .285. A single raises the win
probability to .816, an increase of .101. An out lowers the win probability to
.637, a decrease of .078. The average change (ignoring whether the change is
positive or negative) is .1(.285)+.2(.101)+.7(.078) = .1033. Tango estimates
the average change in win probability during the course of a game to be .0347,
so our situation is .1033
.0347 ≈ 3 times as important as a normal play. The leverage
index of 3 quantifies the significance of a situation, and is useful in quantifying
clutch play.

Game Control and the Story Stat

On May 19, 2015, the Minnesota Twins beat the Pittsburgh Pirates 8-5
and the Washington Nationals beat the New York Yankees 8-6. Based on the
scores, the games seem to have been similar. In fact, the Nationals rallied
from a 6-2 deficit and won on a walk-off home run. The Twins took an early
8-1 lead and cruised to an easy win, giving up meaningless single runs in the
fourth through seventh innings. We can quantify the closeness of the games
using win probabilities.
At baseball-reference.com, the play-by-play game score includes win prob-
abilities after each at bat. Danny Santana opened the game for Minnesota
by grounding out; this lowered the Twins’ chances from 50% to 48%. Brian
Dozier homered to give the Twins a 1-0 lead, raising the win probability to
59%. In the second inning, Joe Mauer hit a bases loaded double to make the
score 5-0, with a win probability of 90%. The win probability never dropped
below 83% the rest of the game. The average of these win probabilities gives
an idea of the closeness of the game. For the Twins, add 48 and 59 and so on
and divide by the total number of at bats. The result is 89.1, showing that
for most of the game the Twins were in a commanding position. By contrast,
the average for the Nationals-Yankees game, a back and forth affair, is 52.3,
showing that the Nationals were not in a strong position for much of the game.
 Sports Strategies 189

The graphs of the win probabilities for the two games gives a visual of the
difference in the games, belying the similar final scores. Baseball Prospectus
calls this type of graphic the “Story Stat” because of the immediate visual
evidence of the ebbs and flows of the game (or, in the case of the Twins-Pirates
game, the lack thereof). The sudden impact of the Nationals’ walk-off home
run is immediately apparent.

FIGURE 9.2: Win Probabilities for Twins and Nationals

The selection committee for the 2015 college football playoff received some
unwanted publicity for using average win probabilities, which they called
“game control,” as a measure of teams’ strengths. Florida State, which had
fallen behind by double-digit points in several games before winning, was par-
ticularly unhappy at having this aspect of its play quantified.

Game Theory
Similarities between baseball pitching and tennis serving include launching
projectiles at high speeds at opponents who have less than a half-second to
judge the trajectory of the ball and make contact with it. Pitchers and servers
vary the speed, spin, and location of their missiles to keep their opponents
off-balance and guessing. The mathematical analysis of this type of cat-and-
mouse game is called game theory.
Game theory gained some cultural popularity from the movie A Beautiful
Mind about one of game theory’s pioneers, John Nash. The name “game the-
ory” does not send people the right message. James Case titled his book on
game theory Competition, and competition theory is a much more accurate
name. Game theory looks at “zero-sum games” in which whatever happens
positively for one competitor happens negatively to the other, and “non-zero-
sum games” in which cooperation can be a positive. The most direct applica-
tions to sports are the zero-sum games.
Penalty kicks in soccer are well modeled by game theory. The kicker has
options on where to kick the ball. The goalie does not have time to see where
the ball is going and react, and so must choose a direction to move. To show one
190 Sports Math

possibility in game theory, consider a young player who kicks right-footed. His
or her kicks are stronger and more accurate when aimed to the left. If the goalie
guesses correctly that the ball is going to the left, the kick is successful 60% of
the time. If the goalie guesses incorrectly, the kick to the left is successful 100%
of the time. However, if the kicker goes to the right, the kick is successful only
20% of the time when the goalie guesses correctly and 40% when the goalie
guesses incorrectly. We summarize this information in, as you might guess, a
matrix that is called the payoff matrix of outcomes.

Kicker
Left Right
Goalie Left 60 40
Right 100 20

For this payoff matrix, the label “Left” for the goalie means that the goalie
is guessing that the kicker will kick the ball to the left (which is the goalie’s
right).

Example 9.6 Determine

 the
 best strategies for the kicker and goalie with
60 40
the payoff matrix
100 20
Solution Mathematically, we analyze this game in the following way. From
the kicker’s perspective, whether the goalie guesses left or right, the percentage
is higher if the ball is kicked to the left. The left column dominates the right
column, in that each entry is larger. Therefore, the proper strategy is a “pure
strategy” of always kicking to the left. From the goalie’s perspective, the worst
that can happen guessing left is for the kicker to go left and score 60% of the
time. The worst that can happen guessing right is for the kicker to go left and
score 100% of the time. The best of the worst (this is called the “minimax”
since we are choosing the smaller of the two larger numbers) is to guess left,
which is the goalie’s pure strategy.

In sports terms, we come to the same conclusion. The kicker gets a better
result going to the left, no matter what the goalie does, so the kicker should
always go left. The goalie, giving the kicker credit for making the right choice,
knows that the kicker is going left and so must guess left to minimize the
damage.
Let’s make Example 9.6 more realistic by changing the payoff matrix to

Kicker
Left Right
Goalie Left 64 89
Right 94 44
 Sports Strategies 191

The kicker is still stronger going to the left, being successful 64% of the
time when the goalie guesses correctly. When the kicker goes right and the
goalie guesses correctly, the success rate is only 44%. However, if the goalie
guesses left, the kicker is better off fooling the goalie by going to the right. It
is now a guessing game, although an interesting one because the kicker would
prefer going to the left. But the goalie knows that and so will tend to guess
left, which would make the kicker go to the right, ..., so that we have quite a
logical web to untangle.
The first main point to make here is that the players do not have a pure
strategy to use. There is not a dominant row or column. Further, either player
can take advantage of knowledge of the other player’s strategy. That’s our
logical tangle: if the goalie knows the kicker is going left, the goalie will go
left, but if the kicker knows the goalie is going left, the kicker will switch to
going right, and so on.
This leads to the second main point: the optimal strategy is a “mixed
strategy” in which each player uses each of the available strategies a fraction
of the time. The challenge is to figure out the best fractions: should the kicker
go left 50% of the time? 60%? How often should the goalie guess left?
The basic principle for determining the optimal percentages is that a given
player does not want the other player to be able to gain an advantage by
knowing the strategy. Both players, in theory, should be able to announce
their percentages and know that the other player cannot use the information
to improve the odds. Example 9.7 shows how this can be done.

Example 9.7 Determine

 the best strategies for the kicker and goalie with
64 89
the payoff matrix .
94 44
Solution Start by giving names to the variables we want to find. Let k be
the fraction of time the kicker goes left and g the fraction of time the goalie
guesses left. Then the kicker goes right 1 − k percent of the time and the
goalie guesses right 1 − g percent of the time. (If that is not clear, think of
an example. If the kicker goes left 40% of the time, then the kicker goes right
60% of the time, and we have fractions k = .4 and .6 = 1 − k.) Replacing the
strategy names with the probability names will give us a handy reference.

Kicker
k 1−k
Goalie g 64 89
1−g 94 44

First, compute the expected outcome if the kicker goes left. The kicker
going left is the first column of the payoff matrix, so look at the two numbers in
192 Sports Math

the first column and the corresponding probabilities to their left. The outcome
64 occurs with probability g and the outcome 94 occurs with probability 1−g.
The expected value is 64g + 94(1 − g). Next, compute the expected outcome
if the kicker goes right. This is the second column of the payoff matrix, so
match the numbers in the second column with the probabilities to the left.
The outcome 89 occurs with probability g and the outcome 44 occurs with
probability 1 − g. The expected value is 89g + 44(1 − g). The basic principle is
that these need to be the same, so we set them equal to each other and solve.
From 64g + 94(1 − g) = 89g + 44(1 − g) we get 64g + 94 − 94g = 89g + 44 − 44g
or 50 = 75g and hence g = 50 2
75 = 3 . The goalie should guess left two-thirds of
the time and right one-third of the time.

Switch roles and compute the expected outcome if the goalie guesses left.
This is the first row of the payoff matrix, so match the numbers in the first row
with the corresponding probabilities above them. The outcome 64 occurs with
probability k and the outcome 89 occurs with probability 1 − k. The expected
value is 64k + 89(1 − k). If the goalie guesses right, we use numbers from the
second row of the payoff matrix and the probabilities above. The outcome
94 occurs with probability k and the outcome 44 occurs with probability
1 − k. The expected value is 94k + 44(1 − k). The basic principle is that these
need to be the same, so we set them equal to each other and solve. From
64k + 89(1 − k) = 94k + 44(1 − k) we get 64k + 89 − 89k = 94k + 44 − 44k or
45 = 75k and hence k = 45 3
75 = 5 . The kicker should go left three-fifths (60%)
of the time and right two-fifths (40%) of the time.

If both kicker and goalie follow the optimal strategies, the fraction of suc-
cessful penalty kicks can be found by substituting in g or k into one of the orig-
inal expected values. For example, setting k = .6 in the formula 64k +89(1−k)
gives an expected value of .74: the kicker converts 74% of the kicks.
Example 9.7 is still a major simplification of the penalty kick situation.
A study reported on by Oliver and Wilson divides the net into nine regions
with left-middle-right horizontally matched with low-middle-high vertically.
The results are interesting. Going high is high risk/reward, with several shots
going too high but with no goalie saves on high shots in the study. Over 70% of
the saves occurred on low kicks. A study reported at scienceofsocceronline.com
indicates that the goalie dives left or right on 94% of shots. Kickers going down
the middle were successful 87% of the time, compared to 83% overall success.
A third study published by Chiappori, Levitt, and Groseclose looked only
at horizontal placement, divided into left-middle-right. They noted that save
rates were nearly equal when the goalie guessed left or middle or right. The
actual percentages are not close to optimal, but having equal save percentages
is a property of the optimal solution and indicates how athletes often find
optimal solutions. If the save percentage was low in one area, goalies would go
 Sports Strategies 193

in that direction less often, making their left/middle/right breakdown closer

to the ideal.

Upsetting the Game Theory

The basic philosophy of game theory is very conservative. Strategies min-
imize the worst outcome that can occur, and guard against knowledge of the
strategy being exploitable. This does not mean that game theory recommends
conservative strategies.
In football, there has long been a theory that to win, a team must establish
the run. Indeed, in 2014 the correlation between rushing yards and wins in
the NFL is a solid 0.4. Recall that correlation does not imply causation; in
other words, correlation shows that winning teams tend to have high rushing
totals, but it does not mean that a team will win more if they run more. In
fact, the 2014 correlation between rushing attempts and wins is a puny 0.05.
The modern football offense has a mixture of runs and passes, and it is
sometimes said that a team must establish the run to be able to pass. This is
partially in line with game theory: mixed strategies should be used randomly
so that the opponent has no idea what is to happen next. The full game theory
lesson is that you must establish both run and pass to be able to run and pass.
The opponent needs to know that you might run up the middle or throw a
deep pass.
There are times when the measurement criterion is not expected score but
probability of winning. There may be some games where losing by 6 is much
better than losing by 30, but there may also be games where all you care
about is winning. Then you might want to move away from the game theory
strategies and try something different. The conservative long-term strategies
of game theory are great for the better team, but how can the underdog team
pull off an upset?
The concept of a high-variance strategy is important. To put this in bas-
ketball terms, suppose your team is a 10-point underdog, with mean scores
of 60 for and 70 against. If each score has a standard deviation of 3, then
the ranges of likely scores are 54-66 and 64-76, with almost no overlap for an
upset to occur. If, instead, each score has a standard deviation of 6, then the
ranges of likely scores are 48-72 and 58-82. There is now a much larger range
of overlap and a better chance of an upset.
Peter Keating and Jordan Brenner of ESPN The Magazine have a formula
for identifying possible upsets that relies heavily on high-variance character-
istics such as pressing defenses, reliance on three-point shooting, and other
risky strategies. Interestingly, in recent years several teams that would have
been likely “giant killers” in the NCAA Tournament were themselves upset
194 Sports Math

in their conference tournaments and did not reach the NCAA Tournament.
High-variance strategies cut both ways.

Getting and Giving Two

A basketball team gets the ball late in a quarter, with about 15 seconds
more on the game clock than on the shot clock. The television announcers
immediately start talking about going “two-for-one.” This strategy calls for a
quick shot, so that there is time to get the ball back a second time at the end
of the quarter. Without the quick shot, the other team gets the last shot of the
quarter. For example, in the NBA with its 24-second shot clock, taking a quick
shot with 30 seconds left in the quarter means that you will get the ball back
with at least 6 seconds left (assuming no turnovers or offensive rebounds). If
you don’t shoot until 20 seconds are left, you might not get the ball back.
A quick calculation makes the two-for-one look like the right strategy.
Suppose that you average 1 point per possession. Taking two quick shots
might lower that efficiency by 20% to 0.8 points per possession. Nevertheless,
you can expect 1.6 points on your two quick possessions as opposed to 1 point
for one long possession. Reality makes this far more complicated. If you take
your time and shoot with 20 seconds left, the other team will often take a
quick shot and leave you 4 or 5 seconds for the last shot. Offensive rebounds
and turnovers occur with some regularity, as well.
In The House Advantage, Jeffery Ma cuts through this theoretical knot
with data. For a given time remaining in the quarter, he computed the average
score for the rest of the quarter. With 45 to 120 seconds left, the offense has
a 0.5-point advantage. With 3 to 20 seconds left, the offense has a full point
advantage. The smallest advantage is 0.25 points with 33 seconds. This is an
ideal time to shoot, when a change of possession is least costly.
Another end-of-game situation that creates debates is when a team is
ahead by 3 points with time running out. Should they let the other team
attempt a three-point shot to tie the game, or foul them and make them
shoot two free throws? Again, a quick calculation gives a decisive answer that
is not necessarily correct. Suppose the other team makes 35% of its three-
point attempts. The game gets tied 35% of the time if they are allowed to
shoot the three. If fouled, the other team needs to make a free throw (let’s
say 80%), intentionally miss the next free throw but get the rebound (maybe
60%) and then make a shot to tie (maybe 50%). Multiplying the probabilities,
the team has a probability of .8x.6x.5 = .24 of accomplishing the full chain of
events, much smaller than the 35% chance. Therefore, foul them. However, if
you foul them with 6 seconds left, the other team has time to foul you right
back and then get the ball back with 5 or so seconds left, plenty of time to
get a good shot (and, if you missed one of your free throws, they don’t have
 Sports Strategies 195

to shoot a three this time). Plus, if they shoot and make a three-pointer with
6 seconds left, then you have a chance to win the game with a last shot. The
final decision is not clear.

The Physical Challenge

Not all issues of strategies in sports are easily quantified and analyzed. A
couple of interesting situations are discussed here.
In the 2008 Olympics, Michael Phelps came from behind in the 100 m
butterfly to win by .01 second (the limit of the timing at the Olympics) over
Milorad Cavic. Phelps took what most coaches consider an ill-advised last
stroke before hitting the wall, while Cavic coasted into the wall. Phelps’s
stroke was perfectly timed and his fingertips hit the wall just before Cavic’s.
A similar situation occurs frequently in baseball when a runner tries to
beat a throw to first base by diving to the bag. Is it better to dive than to run
all the way through the bag? It is mostly an issue of timing. Experiments tend
to find diving slower, but it depends on whether the runner dives onto the bag
(fast, if dangerous for jamming a finger) or hits the ground first and slides into
the bag (slow). As for running, if the runner’s natural rhythm brings a foot
down onto the bag, that is faster. If the runner must shuffle his feet to hit the
bag, that is much slower.
Reaction times were discussed briefly in previous chapters. In a race, is it
worthwhile trying to anticipate the start of the race to get a slight edge on the
competition? In swimming and track, the penalty for a false start (starting
early) is typically disqualification. On the track, pressure plates in the starting
blocks can determine when a sprinter begins to run, and anyone who starts
less than 0.1 s after the gun fires is deemed to have false-started. This is based
on extensive research into optimal reaction times to auditory signals, which
indicates that a reaction time of 0.12 s is near the limit of human abilities.
Reaction time should be positively correlated with overall time: the sooner
you start, the sooner you finish. However, data published by the IAAF con-
sistently shows little or no correlation between reaction time and overall time.
The explanation could be that you can become an elite sprinter by having a
fast start or by having a fast finish, and the long legs that can help you run
fast will slow down your start.
Reaction time is different in drag racing. The timing mechanism is precise
and predictable, a series of lights that flash at regular intervals. Drivers can
train themselves to start at exactly the right time. An episode of Sports Science
clocks Hillary Will with a reaction time of 0.001 s. Clearly, she is not reacting
but instead is anticipating. Precision timing is obviously a skill that drag
racers need.
196 Sports Math

Personnel Decisions: Aging

Some of the most important analytics work being done supports personnel
decisions. Much of this is proprietary, complicated, or not that interesting to
the common fan. A brief discussion follows on three aspects of this area of
analytics.
The concept of a peak age is simple enough to understand. Young athletes
have not fully honed their craft, while old players are fighting against physical
decay. Somewhere in the middle is the age at which the best performances
occur. The question is how to determine this peak age. But, let’s be more
precise, more mathematical. What exactly are we measuring? The peak age
for batting average could be different for the peak age for home runs, and so
on. You could look at a specific skill or some measure like WAR of overall
performance.
Having decided on a statistic to track
(I’ll use home runs as an example), we
next decide what data to collect. We
could pick a year like 2014 and find how
many home runs were hit by 20-year-
olds, then 21-year-olds, and so on. Figure
9.3 shows the result. We could conclude
from this that the peak age for home
runs is 28. But, there might be more
home runs hit by 28-year-olds because FIGURE 9.3: Home Runs
there are more 28-year-olds playing. The
large drop in home runs at age 29 is more likely due to fewer 29-year-olds in
the league than to 29 being a jinxed age. So, we should control for that by
computing home runs per 500 at bats in each age group.
Figure 9.4 does this, and the result is
quite different. Other than a rise in home
run rate in the late 30s, the rates look
fairly equal at all ages. Looking carefully
at ages 23-28, you can see a general up-
ward trend that is maintained. The lack
of a clear downward trend following age
28 could be due to small sample sizes.
You could reasonably object that Figure
9.4 only provides a snap shot of what
FIGURE 9.4: HR Rate
happened in 2014. If we repeat the fig-
ure for 2015, will we find that 29-year-olds are the best? Perhaps this graph
is more a function of selection bias (which players happen to be grouped to-
gether) than how players of all abilities age.
Let’s look at a specific player. Ian Kinsler had his first full season in the
 Sports Strategies 197

major leagues at age 24, hitting 14 home runs. In succeeding years, his totals
were 20, 18, 31, 9, 32, 19, 13, and 17. Kinsler only played 103 games in the
year in which he hit 9 home runs, so a good analysis needs to account for
injuries. He changed teams in 2014, so it would be good to account for the
effects of ball parks. However, he hit the most home runs at ages 27 and 29,
and there is a general drop before and after those years.
With this in mind, we could select a group of players with long careers
and track their ups and downs. The selection bias here is that we are only
looking at very good players, good enough to stay in the league for many
years. Perhaps this group ages more gracefully than the average player. Still,
this seems like a good way to go. Bradbury did a study like this and found
peak ages of 28 for batting average, 32 for walks, 30 for home runs, and 29
for Linear Weights (see Example 7.5). Others have found a peak baseball age
of 27, but in reading studies on this topic remember that different aspects of
a sport may have different peak ages.

Personnel Decisions: Transfer Fees and Stars

A second issue with personnel implications is transferability of skills. The
underlying issues here are individual skill and “fit” in a particular team’s
system. For example, if a running back ran for 1800 yards with one team and
then changes teams, is it reasonable to expect another 1800-yard season? The
book The Success Equation says no, that one of the few skills that holds up
after changes of teams is punting.
How would you investigate this? Similar to aging, you need to choose a
particular statistic to track. As far as collecting data, a simple idea would be
to collect all before-and-after team changes for that stat. There are several
factors you might adjust for. Style of the team is important: our running back
moving to a pass-heavy team would not have the chance to duplicate his 1800-
yard season. You could use yards per carry or percent of the team’s rushing
yards instead of total yards. In the year before becoming a free agent, players
tend to have productive years, and then drop off after signing. Part of the
drop-off is regression to the mean; our 1800-yard running back was likely to
drop back closer to the league average if he stayed on his original team.
A third personnel issue has to do with team composition. In Chapter 8, we
see that basketball teams with unequal scoring distributions tend to perform
well. The theory that you need multiple stars to win in basketball is explored
in the book Scorecasting. The short answer is “yes you do.”
Anderson looks at this issue for soccer in The Numbers Game. As his
measurable statistic (metric), he uses the Castrol Performance Index, which
(like WAR for baseball) uses play-by-play data to try to measure the total
contribution of a player to a team. He then looks at how teams’ records depend
198 Sports Math

on the value of the top players on the teams and on the bottom players on
the teams. Not surprisingly, a team’s record improves as its star player gets
better, and also improves if its worst player gets better. The interesting result
is that the value of the worst player has more of an effect on overall team
performance than does the value of the star player. In this way, soccer seems
to be a team sport in which the chain is only as strong as its weakest link.
To try a similar study, I took the eight position
starters for each 2014 MLB team and computed
the sums of the three highest and three lowest
WAR values, as posted on baseball-reference.com.
Figure 9.5 shows the scatter plot of the best WAR
values against team wins. There is a mild trend
for better teams to have higher top-three WAR FIGURE 9.5: Best 3
values; the correlation is 0.53.
On the other hand, Figure 9.6 shows the scat-
ter plot of the worst three WAR values against
team wins. There is little difference visually be-
tween Figures 9.5 and 9.6, and the correlations of
0.53 and 0.49 confirm that the worst WAR values
predict team wins as well as the best WAR values
do. FIGURE 9.6: Worst 3
A different version of the same study is to do a
multiple regression, which produces the best-fit linear function 68.10 + 6.17b +
2.96w for wins, where b is the WAR value of the best player and w is the
WAR value of the worst player. Since the coefficient of b is larger than w, we
conclude that an increase in WAR value of the best player is more important
than an increase in WAR value of the worst player. The “r-squared” value of
the model is 0.49, indicating that 49% of the variation in wins can be explained
by variations in WAR values (and that the model and wins have a correlation
of 0.7). Pitchers were not included in the data. This gives some mild evidence
that baseball is a star-oriented sport.
The conclusions here are dependent on the data that go into the study. If
the player ratings are flawed, then any conclusions will be flawed. However, the
Oakland A’s and other low-budget sports teams have been using a “bottom-
up” strategy of team building. They can’t afford the big stars with the huge
salaries, but they can build a strong team by making sure that even the worst
players on the team are good. Getting a few young players to perform at high
levels is their hope for gaining the boost that stars provide.
 Sports Strategies 199

Exercises
9.1 Using p(x) = 6 − 0.07x, estimate the expected number of points for a football
team starting at (a) its own 20 (b) its own 10. (c) At what position is the
expectation 0? (d) Explain why this might be a fair place to start an overtime
period.
9.2 Compute expected score values of going for it versus punting (assume a 34-
yard net gain) on 4th and 2 at (a) your own 20 (b) your own 10. (c) How crazy
is Kevin Kelley to never punt?
9.3 In example 9.2, the probability of making the first down in 0.53. How small
would the probability have to be for the punt to have the larger expected value?
9.4 Suppose the result of an onside kick is for one team or the other to recover at
the kicking team’s 45. Find the break-even point, the probability of recovering
the kick such that the onside kick has a positive expected value.
9.5 Suppose the result of a kickoff is for the receiving team to get the ball at its
35. Find the break-even point of an onside kick (see exercise 9.4), the probability
of recovering the kick such that the onside kick has a larger expected value than
a kickoff.
9.6 Suppose the options on an extra point are a sure 1 point or 2 points with
probability 0.53. Which has the larger expected value?
9.7 Find the points value of a pass play from a team’s own 20 to the 50. Show
that the value is the same for a 30-yard gain from any location.
9.8 For the OneBase model, change the probabilities to 0.2 for a homer, 0.3 for a
hit, and 0.5 for an out. (a) Write the transition matrix. (b) Write the expected
runs vector for one batter. (c) Write the equations for total expected runs. (d)
Solve those equations.
9.9 Find, as in Example 9.4, the expected run values for a runner on first with 1
out versus a runner on second with 2 outs,
9.10 Repeat example 9.5 with 1 out.
9.11 Suppose a tennis player wins points with probability 0.6 if the score is tied,
0.7 if ahead, and 0.4 if behind. If a game is won by the first player to win two
points, set up the transition matrix for the states A (0-0), B (1-0), C (0-1), D
(1-1), E (game won), F (game lost).
9.12 Given the win probabilities (half-inning averages, from baseball-
reference.com) from August 3, 2015, compute the average win probabilities. (a)
Atlanta 9, San Francisco 8: 54, 52, 30, 31, 21, 12, 5, 5, 3, 3, 2, 14, 19, 28, 27,18,
13, 50, 55, 100 (b) Texas 12, Houston 9: 39, 74, 83, 83, 83, 96, 94, 91, 93, 95, 96,
97, 96, 94, 97, 97, 100
9.13 A simple game for pitcher versus batter has a pitcher throwing a fastball
or slider. When the batter guesses correctly, he bats .400 against the fastball
and .300 against the slider. When the batter guesses incorrectly, he bats .200
against the slider and .350 against the fastball. Determine the best strategy for
the pitcher and batter and the resulting batting average.
9.14 Repeat exercise 9.13 changing the .350 value to a more realistic .250.
200 Sports Math

9.15 The table gives minutes played and points scored by age in the 2014-15 NBA
season. (a) Which age scored the most points? (b) Which age scored the most
points per minute? (c) Discuss the optimal age for scoring points in the NBA.
20 21 22 23 24 25 26 27
points 4533 11199 15984 19173 24540 23045 25335 18627
minutes 13646 28898 37387 50090 55979 51958 60456 46161
28 29 30 31 32 33 34
points 21864 21914 13716 6817 8726 7275 8926
minutes 53652 51822 32524 17045 21826 17515 25154
9.16 In the 2014 American League season, the numbers of wins were {96, 71, 73,
85, 90, 70, 89, 98, 70, 84, 88, 87, 77, 67, 83}, the highest WAR values by team
were {6, 4.8, 5.5, 7, 5.5, 6, 6.6, 7.9, 5.2, 4, 7.4, 6.4, 5, 7, 6} and the lowest WAR
values were {.7, -1.4, -1.1, -1.4, -1.5, -.8, -1.2, -.3, -1, -1.4, -1, -1.5, -2, -1.9, -.5}.
Create scatter plots and compute correlations of WAR values to wins.
9.17 In the 2014 American League season, the numbers of wins were {96, 71, 73,
85, 90, 70, 89, 98, 70, 84, 88, 87, 77, 67, 83}, the sums of the WAR values for the
starters were {32.2, 19.4, 16.5, 18.1, 23.4, 14.1, 21.4, 33.2, 20.5, 13.3, 26.0, 19.7,
19.2, 14.9, 24.4} and the sums of the WAR values for the subs were {6.0, 2.0,
-1.6, -1.6, 3.5, -0.2, -0.6, 2.3, 0.6, -0.5, 3.3, -0.7, -0.5, -2.4, 2.8}. Create scatter
plots and compute correlations of WAR values to wins.
9.18 T (a) Discuss circumstances in which win probability added is a more
informative football statistic than expected points added. (b) For evaluating the
effect of a team’s offense, discuss circumstances in which win probability added
could unfairly treat mediocre and great teams differently.
9.19 T Discuss, in terms of expected points, whether a fumble is equally harmful
at any location on the field. Give examples where win probability would change
dramatically for a fumble at one location but not another.
9.20 T Discuss ways to implement a truly random strategy.
9.21 T Discuss why soccer goalies almost always dive left or right for penalty
kicks, instead of playing the middle as often as calculations suggest is ideal.
9.22 T Discuss why penalty kick takers do not go high very often, even though
high shots are almost never saved.
9.23 T Discuss how a team (sport of your choice) might implement a high-
variance strategy. Does Kevin Kelley’s strategy qualify?
9.24 T Discuss whether NBA teams who use the 2-for-1 strategy take lower-
percentage shots.
9.25 T Discuss the differences between baseball, basketball, and soccer in terms
of whether the quality of the best player is more important than the quality of
the worst player.
9.26 T In Analyzing Wimbledon, it is reported that inexperienced players tend
to hit conservative serves in pressure situations, whereas the top players show
little or no change in service speed and placement. In game theory terms, discuss
why it is important to maintain the same strategy in pressure situations.
9.27 T In The 1984 Baseball Abstract, Bill James notes that the 1983 National
League East was highly compressed, meaning that the difference between first
and last place was small. He also noted that the Cubs won fewer games in 1983
 Sports Strategies 201

than their Pythagorean Method projection said they should have won. For these
and other reasons, James called the Cubs a good long-shot bet, a team that
could go from last to first. Explain why this was a reasonable prediction. (And,
it turns out a good one: the 1984 Cubs did win the division.)
9.28 C For the OneBase model, (a) write out the matrix S and (b) compute
(I−S)−1 .
9.29 C Derive the formula (I-S)−1 R for the total expected runs in the OneBase
model.
9.30 C Compute the change in expected runs (use Tom Tango’ expected runs
matrix) for the following plays. (a) a leadoff single (b) a leadoff double (c) a
single with one out and nobody on (d) a double with one out and nobody on (e)
with one out and a runner on first, a single that advances the runner to third (f)
with one out and a runner on first, a single that advances the runner to second
(g) How much does the value of a single change with the number of outs? (h)
How much more is a double worth than a single? (i) How many runs is it worth
for a runner on first to get to third on a single?
9.31 C Find the change in win expectancies for the situations in exercise 9.29
parts (a), (b), (e), and (f) assuming that (1) the score is tied in the bottom of
the ninth (2) the team is down one run in the bottom of the ninth.
9.32 C The probability of winning a service point in tennis is ab + (1 − a)xy,
where a is the fraction of first serves that are in, b is the fraction of points won
when the first serve is in, and x and y are the corresponding values for the second
serve. (a) In Analyzing Wimbledon, x = 0.86 and y = 0.51. If the relationship
between fraction in and fraction won is b = 0.93 − 0.5a, find the optimal first
serve percentage. Compare to the actual average of 60%. (b) If only one serve
is allowed, then the probability is ab. Find the optimal serve percentage. Is this
closer to the actual first serve percentage of 60% or the second serve percentage
of 86%? Discuss.
9.33 C Team A leads by 3 points with 4 seconds left and fouls team B. Team
B then fouls team A with 2 seconds left. Team B launches a 3-point shot at
the buzzer. Assuming 70% free throw shooting, 30% 3-point shooting, and no
offensive rebounds, what is the probability that team B ties the game? wins?
9.34 C Reaction times and final times in the 2013 IAAF World Championship
men’s 100 meter sprint final are given. (a) Explore the importance of reaction
times with a scatter plot and correlation. (b) For a given sprinter, lowering
reaction time from 0.20 to 0.15 would improve the overall time by 0.05 s. Would
that make a difference in placement for anybody? (c) Part (b) argues for a direct
correlation between reaction time and final time. Reconcile this with the small
correlation in part (a).
1st 2nd 3rd 4th 5th 6th 7th 8th
reaction 0.163 0.163 0.157 0.186 0.142 0.158 0.154 0.177
final 9.77 9.85 9.95 9.98 9.98 10.04 10.06 10.21
9.35 C Compare results from pro-football-reference.com’s win probability cal-
culator (use a Vegas line of 0, and 1st and ten) for the following situations. (a)
trail by 1, 3 minutes remaining, 30 yards away; (b) trail by 1, 3 minutes remain-
ing, 20 yards away; (c) trail by 1, 2 minutes remaining, 30 yards away. Then (d)
estimate the value of 10 yards and one minute late in a game trailing by 1.
202 Sports Math

9.36 P Expected points from different starting positions in football are nearly
linear in distance from the goal line for the entire season. Investigate whether
the linear relationship holds in special situations, such as the fourth quarter of
a tight game or for a specific team.
9.37 P Overall, NBA players have a higher shooting percentage in catch-and-
shoot situations than for pull-up shots. Is this true of all teams? all players?
For a potential game-winning shot, is the higher percentage play to go to the
team’s best shooter (who might have to take a dribble to get free) or kick it to
a wide-open lesser shooter for a catch-and-shoot?
9.38 P From game-by-game results, determine the correlation between yards
rushing and winning. Would a high correlation prove that rushing creates wins?
Determine the correlation between rushing yards in the first quarter and wins.
Discuss whether rushing causes victories.

Further Reading
The Sloan Sports Analytics Conference videos are available online. “In-
Game Innovations: Genius or Gimmick” is the panel discussion with Kevin
Kelley and Bill James. Keating and Brenner’s talk is “Giant Killers.” Jeff van
Gundy’s comment is in 2015’s “Innovators and Adopters.”
More on Kevin Kelley can be found at http://grantland.com/features/
grantland-channel-coach-never-punts/ Accessed 8-11-2015.
The books Moneyball and Big Data baseball give extended examples of the
influence of analytics on baseball strategy.
In The Hidden Game of Football, Carroll, Palmer, and Thorn develop the
concept of Win Probability, while exploring many of the ideas in this chapter.
“Monday Morning Math Modeling” by Hodds, Alcock, and Inglis in the
February 2014 issue of the journal Math Horizons explores Belichick’s strategy
in the 2012 Super Bowl.
Expected runs matrices can be found at numerous sites and books, includ-
ing Tom Tango’s book The Book and website. Play-by-play win probabilities
are part of the game descriptions at baseball-reference.com.
Win probabilities for football can be found at pro-football-reference
(www.pro-football-reference.com/play-index/win prob.cgi accessed 8-27-15).
Hot Stove Economics by Bradbury discusses peak ages in baseball.
Anderson and Sally’s book The Numbers Game and Eastaway and
Haigh’s How to Take a Penalty discuss penalty kicks. Articles referenced
are Chiappori, Levitt (yes, the Freakonomics co-author), and Groseclose:
http://pricetheory.uchicago.edu/levitt/Papers/ChiapporiGroseclose
Levitt2002.pdf accessed 8-27-15 and William Spaniel’s http://williamspaniel
.com/2014/06/12/the-game-theory-of-soccer-penalty-kicks/ accessed 8-27-15.
An excellent model of aging in sports is in “Bridging Different Eras in
Sports” by Berry, Reese, and Larkey.
Chapter 10
Big Data and Beyond

Introduction
A baseball batter hits a hard groundball
into the hole. The shortstop, known as one of
the best fielders in the game, backhands the
ball and makes a spectacular jump-throw to
first that the runner barely beats out. The
batter is on first with a hit, but everyone
applauds the all-out effort by the shortstop.
That tale from the low-tech past may well
be replaced, by the time you read this, with
the following description. The batter hits a
ground ball with exit velocity 81 mph. The
shortstop has a very slow reaction time of 0.4
s, and only covers a distance of 8 feet before
fielding the ball. The time between fielding
the ball and getting off the jump throw is
a very slow 1.3 s, and the throw has a low
velocity of 58.3 mph. An average shortstop FIGURE 10.1: Jump
would have thrown out the runner 95.8% of
the time, so the official scorer flashes “E6” indicating an error on the shortstop.
This chapter is about the brave new world of big data in sports. Massive
data sets are coming online for all major sports. The challenge for teams
will be to mine information that gives them competitive advantages. From a
fan’s perspective, this chapter is about the future as seen from 2015. Some of
what is included here will be overly conservative; hopefully, little of it will be
laughably off target. The following questions will be addressed. What types
of new data are now available? What are some ways in which the presence
of this data can help to measure previously unknowable quantities? What are
some ways to visualize data that arrive in quantities that boggle the mind?
Will big data affect the way that the casual athlete participates in sports?

203
204 Sports Math

Big Data Is Watching You

Modern sports contests are data collection extravaganzas. Arenas bristle
with electronics, as cameras and sensors record every action on and off the
field of play. Armies of computers sort and manipulate the data into forms
that are useful for participants, media, and fans.
Through the All-Star break in 2015, Max Scherzer had thrown six different
types of pitches (information at the Brooks Baseball website, from MLBAM
data). His go-to pitch is the four-seam fastball, thrown 57% of the time; he
only throws the two-seam sinker 1% of the time. How can we tell? The trajec-
tories of his pitches are reconstructed by PITCHf/x, and different pitches move
different distances in different directions. Scherzer’s four-seamer averages 7.73
inches of horizontal movement and 7.92 inches of vertical movement, as com-
pared to 9.66 and 7.31 inches, respectively, for his sinker. His average release
point is more than two inches wider for his sinker than for his four-seamer.
He throws his sinker three times as often against right-handed batters.
In 2014-15, Philadelphia’s Michael Carter-Williams led the NBA with
100.5 touches per game, and also with 73.9 passes per game. Jimmy Butler of
Chicago led the league with an average of 2.7 miles run per game. San An-
tonio’s Patty Mills had the highest average speed at 4.8 mph. Among players
who had at least 100 drives on the season, Toronto’s James Johnson had the
highest field goal percentage with 63.7% made. Detroit’s Reggie Jackson led
the league with 15.6 points per game created (scoring and assists) on drives.
Minnesota’s Ricky Rubio led the league with 1.3 free throw assists per game
(passes to players who were fouled, missed the shot, but made at least one
free throw). The Clippers’ Chris Paul had 2.8 second assists (passes that led
to assists) per game, and created 23.8 points per game from assists; both were
league highs. Oklahoma City’s Kevin Durant shot 43.5% on pull-up threes (at
least one dribble), while Atlanta’s Kyle Korver shot 50.4% on catch-and-shoot
threes (no dribbles). (All numbers from nba.com.)
Most major soccer teams record locations of all players several times per
second, from which velocity and acceleration can be computed. Soccer data
can be hard to come by, but from www.dailymail.co.uk we learn that Lewis
Holtby of Fulham ran 8.1 miles in a match against Liverpool on July 15, 2015.
The day before, Jay Rodriguez of Southampton had reached a speed of 21.5
mph playing against Hull City. From the matchcenter at mlssoccer.com we
learn that the New York Red Bulls connected on 48% of its crosses in its U.S.
Open Cup match against the New York Cosmos. The Cosmos only completed
23% of its crosses and lost 4-1.
A graphic on the ESPN telecast of Wimbledon 2015 showed that Roger
Federer never hit a first serve into Andy Murray’s body in their semifinal
match. Other graphics showed the positions on the court where the players
 Big Data and Beyond 205

hit returns of serve, with Federer much farther into the court than Murray.
This data has not been released to the public for analysis.
At the 2015 MIT Sloan Sports Conference, Andrew Hawkins of the Cleve-
land Browns told a story about wide receivers wearing practice shirts wired
to record speed and distance. The more competitive players would pull their
chips out and throw them down the field to try to record the highest speed
for the day. Coaches use practice data to judge whether players are working
out at the proper level to prepare for the game while preventing injuries.
There is very little that an athlete does now that is not recorded and
analyzed. The challenge is mining the data for useful information that can be
clearly communicated to the coaches and athletes.

Example 10.1 Following up on Examples 9.4 and 9.5, discuss how data
such as pitch movement, hitting charts, and running speeds could change the
optimal strategy.
Solution Many responses are possible, including the following. The conclu-
sion in Examples 9.4 and 9.5 is that the sacrifice bunt is an ineffective play.
However, suppose that the pitcher throws a two-seam sinker that causes bat-
ters to hit ground balls 60% of the time. The batter hits ground balls 55%
of the time, typically pulling them to one of two zones that the defense has
covered by shifting its alignment. The chance of a double play is high, so the
bunt might be a good play. On the other hand, suppose the pitcher throws a
four-seam fastball and tends to give up fly balls. The batter hits line drives
to all fields and has good speed. The chance of a double play is low, and the
chance of a winning extra base hit is high, so the bunt would be a poor play.

A Theory of Everything
The Holy Grail of sports analytics has been a single number that measures
the performance of every athlete in the sport. This formula would identify
the best in the sport, would quantify how much money an athlete is worth in
salary, and would quantify the fairness of a trade or the quality of a draft pick.
A realistic appraisal of these goals should make you question whether such a
formula could possibly exist. Nevertheless, the NFL Quarterback Rating and
baseball’s WAR and other rating systems are popular with the media and
fans, if not the teams.
Baseball’s WAR, soccer’s Castrol Performance Index, football’s Total
QBR, and golf’s Strokes Gained all share the same conceptual framework,
which can be called “Value Added.” First, organize the data so that the av-
erage outcome from any situation can be identified. Then, for every play for
every athlete, compare the athlete’s performance to the average performance,
measured in units that are meaningful for the sport in question.
206 Sports Math

If a leftfielder hits 12 home runs in 132 games, compare that to the home
runs you would expect his replacement to hit and quantify it in terms of wins
above or below replacement level. Do the same for every fielding play and
baserunning situation in the season, and you have WAR.
The Castrol Performance Index assigns points for passing accuracy, shoot-
ing accuracy, fouls given and taken, and so on, weighted based on the location
of the play and how likely the play is to lead to or take away from a scoring
chance. Total QBR takes into account the quarterback’s ability to run, avoid
sacks, deliver the ball to his wide receivers (even if the receiver then drops
the ball), and so on, weighted by the importance of the play as determined by
location on the field, score, and time remaining. In golf, Strokes Gained for a
shot equals the difference between the expected score from the location before
a shot and the expected score from the new location after the shot. Each type
of stroke (putting, driving, and so on) can be evaluated separately or added
together to get a total rating.
The different versions of WAR or Strokes Gained that you may read about
differ in implementation of the system, but work from the same concept. The
definition of a replacement player, the method for measuring the importance
of a play, the granularity of the measurements (should a 92 mph line drive
to left be measured against the same standard as a 93 mph line drive to
left-center?), and the data used can all affect the final rating.

Example 10.2 A golfer plays a 480-yard par 4. He hits a 320-yard drive in

the fairway, hits a 160-yard approach shot 22 feet from the pin, and two-putts
for par. Determine Strokes Gained for each shot on this hole.
Solution A solution depends on several linear regressions on golf data,
adapted from my Golf By the Numbers. From the approximation f (x) =
2.77+.0028x, estimate the average score on a 480-yard par 4 as f (480) = 4.114.
From the approximation g(x) = 3.28+.004x, estimate the average score hitting
the second shot from 160 yards away in the fairway as g(160) = 3.92. From
the approximation h(x) = 3.68 + .01x, estimate the average score putting
for birdie from 22 feet as h(22) = 3.9. The drive lowered the expected score
from 4.114 to 3.92, so the drive saved 0.194 strokes. The approach shot saved
3.92−3.9 = 0.02 strokes and the two-putt cost him 4−3.9 = 0.1 strokes. With
Strokes Gained values of −.194, −.02, and .1 (adopting the convention that
better than average is under par and therefore negative), the player’s overall
rating is −.114, indicating that his 4 is .114 strokes better than the average
(par) of 4.114.

Other examples of Value Added will be presented below.

 Big Data and Beyond 207

Catch Me If You Can

While Big Data can help fine tune overall rating systems, the potential
for changing the way sports are played and watched comes primarily from
shedding light on aspects of the game that were previously unknowable. Anal-
yses of the ability of baseball catchers to handle pitchers and control base
runners have long been limited to vague quotes from players. Percentage of
bases stolen has been used to evaluate catchers, even though everybody knew
that pitchers have a large role in preventing or enabling runners to get good
jumps. (In the three seasons from 2006 to 2008, only 10 runners tried to steal
against Kenny Rogers; 8 were thrown out. Over a six year period Chris Young
was run on 133 times with 126 bases stolen!) In all, we have been woefully
ignorant of the so-called “tools of ignorance” employed by catchers.
Baseball Info Solutions (BIS) publishes The Fielding Bible. The 2015 edi-
tion shows how far we have come. BIS has defined a suite of useful measures of
catching proficiency. On an attempted stolen base, pop time is the elapsed
time from the ball hitting the catcher’s mitt to the ball arriving at the fielder’s
glove for a tag. Then tag time is the elapsed time until the tag is made. Pop
times generally range from 1.8 to 2.2 seconds. The percentage of base runners
thrown out increases almost linearly from 36% for a pop time of 1.8 s to 25%
for a pop time of 2.2 s with a right-handed pitcher on the mound (stealing
quantity and quality are lower if the pitcher is lefthanded).
Tag time is used as a proxy for the catcher’s throwing accuracy. Tag
times for catchers range from 0.20 to 0.24 seconds, enough to matter but
not as important as pop time. As noted, the pitcher plays a large role in
base-running success. As one indicator of this, average caught stealing rates
drop from 30% to 8% as the pitcher’s delivery time increases from 1.3 to 1.65
seconds. And, of course, the speed of the runner is important. Caught stealing
rates increase from 10% to 45% as the runner’s time increases from 3.3 to 3.8
seconds. Additionally, the pitch types (caught stealing averages for fastball are
26-27%, versus 16% for curves) and pitch locations have measurable effects.
There are other factors to be measured, such as ability to hold a runner close
to first base, but the secrets of base-stealing are rapidly being revealed.

Example 10.3 Russell Martin had an average pop time in 2014 of 1.88 s.
Estimate his caught stealing percentage. If the pitcher’s delivery time is 1.4 s
and the runner’s time is 3.5 s, estimate the runner’s chance of stealing a base.
Solution If caught stealing percentage is a linear function with f (1.8) = 36
36−25
and f (2.2) = 25, then f (x) = 36− 2.2−1.8 (x−1.8) or f (x) = 85.5−27.5x. Based
on pop time, Martin’s caught stealing percentage should be f (1.88) = 33.8.
(His actual percentage in 2014 was 32%.) For delivery time, the line through
(1.3,30) and (1.65,8) is approximately g(x) = 111.7 − 62.86x so a delivery
time of 1.4 s should produce a percentage of g(1.4) = 23.7. For running time,
208 Sports Math

the line through (3.3,10) and (3.8,45) is h(x) = 70x − 221 so a time of 3.5 s
should produce a percentage of h(3.5) = 24%. Our three data points produce
estimates of 33.8%, 23.7%, and 24%. While it is not clear how to average them
(more research is needed!), a simple mean is 27.2%.

The next step here is to convert pop and tag times into runs saved (and,
eventually, wins) by the catcher. A simple approach would be to create a
formula into which you can plug in pop time and tag time and get a number.
Given the available data, a better way of evaluating the catcher’s performance
is to do it play-by-play, taking into account delivery time, runner’s speed, and
whatever other information can be gathered. One issue is how many of these
variables to include. In Example 10.3, taking the pitcher’s delivery time and
the runner’s time into consideration, you could reasonably assign an expected
percentage of 25.6% to this situation. If Martin throws out the runner, he gets
credit for 74.4% of an out. Using the expected runs table discussed in Chapter
9, Martin could be given 74.4% of the difference in expected runs before the
play and after the play.

Getting Framed
In the 1900s, fans would sometimes hear vaguely mystical statements about
catchers being good receivers. That means catching the ball, right? It was hard
to figure what the skill might be. Gradually, the skill of “framing” a pitch was
clarified, and in the 2010s several researchers started quantifying the effect.
The idea of framing is that a catcher who lunges for a pitch sends a
signal to the umpire that the pitch was not thrown well, and is probably
outside of the strike zone. The same pitch caught smoothly without the lunge
(its location clearly “framed” by the catcher’s body and glove) is more likely
to be called a strike. That seems reasonable, but the size of the effect was a
big surprise to everyone.
The effects of pitch framing can be measured using the Value Added
method. That is, we can map out the strike zone and record the locations
of all pitches and the resulting umpire calls. A location in the center of the
strike zone might be called a strike 100% of the time, whereas waist high
on the outside corner might be a strike 65% of the time, and knee high on
the inside corner might be a strike 40% of the time. Suppose Russell Martin
catches a pitch in a location that is called a strike 40% of the time, and gets
a strike call. Martin gets credit for +0.6 strikes better than average. Given
data about batting averages for different strike counts, this can be converted
to a fraction of a run saved. BIS’s Fielding Bible states that, on average, the
difference between a ball and strike call is about 0.12 runs. Add up all of the
 Big Data and Beyond 209

runs saved or lost for the entire season and you have an estimate of Martin’s
worth as a receiver.
How many runs in a year would you guess a good catcher saves? The first
estimate (as told in Big Data Baseball, an excellent Moneyball -like chronicle
of the resurrection of the Pittsburgh Pirates) was wild: the difference between
good and bad catchers could be 300 runs in a season! Modern estimates divide
that by a factor of 10, but 30 runs (about 3 wins) in a season is significant.
This estimate does not include the psychological effect on the pitchers of being
ahead in the count or being able to throw a borderline pitch with a full count.
The methodology described above is a first approximation that has already
been adjusted. It makes little sense to give the catcher full credit or blame
for a borderline pitch being called a ball or strike. Umpires have tendencies
for being more or less generous with calls, a hitter who crowds the plate can
affect an umpire’s judgment, and a pitcher with good control is thought to get
more calls than a wild pitcher (a paper by BIS’s Rosales and Spratt finds that
horizontal accuracy is more important than vertical accuracy). Credit and
blame for balls and strikes calls must be divided among the catcher, umpire,
pitcher, and batter.
Big data has revealed that pitch framing is real and an important skill
that can be measured.

Anonymous Field Goal Kicking

The percentage of field goals made in the NFL has steadily risen from
about 42% in 1952 to about 85% in 2012. All fans know how painful it is to
see their kicker miss, particularly near the end of a game. This may disguise
how infrequently NFL kickers miss. You might expect that with misses being
a rarity it is hard to distinguish individual kickers. (Could you tell, without
keeping score, the difference between free throw shooters who make 83 and 85
free throws, respectively, out of 100?) Research by Pasteur and Cunningham-
Rhoads comes to this conclusion.
The structure of this study is to compute Value Added scores for each
kicker. For each kick, compare the result to the expected make percentage for
that type of kick. The word “type” requires definition. Which variables should
be considered? You can argue that a large number of variables affect field goal
percentage: distance, angle, wind, temperature, the quality of the defense,
whether the kicker is tired or not, and so on. At some point, you have to say
that the effect is not large enough to bother with. A linear regression can help
determine which variables are useful and which are not: if the inclusion of a
new variable only improves the total error by 0.01% it is not wise to retain
the variable. In this study, only one of the above variables (angle) did not
210 Sports Math

make the final model, although distance is far and away the best predictor of
accuracy.

1
Example 10.4 Use the formula p = for the probability
1 + e−5.8409+0.1078d
of making a field goal of d yards to compare accuracy from (a) 30 yards, (b)
40 yards, (c) 50 yards.
1 1
Solution. (a) = 0.931. (b) =
1 + e−5.8409+0.1078∗30 1 + e−5.8409+0.1078∗40
1
0.822, only slightly lower. (c) −5.8409+0.1078∗50
= 0.611, still well over
1+e
50%. The trend .931, .822, .611 is not linear, as the drop in accuracy from
40 yards to 50 yards is much larger than the drop in accuracy from 30 to 40
yards.

The distance-only model is reasonably accurate. The study takes advantage

of this by stating the effects of other variables in terms of yards. (However, note
that given the nonlinear nature of the probabilities in Example 10.3, the effect
of an extra yard is much greater at 50 yards than at 40 yards.) The altitude
of Denver subtracts about 3 yards, the difference between kicking against a
good defense and a bad defense (as measured by points per game) can be 4
yards (surprising to me), the difference between hot and cold temperatures
can be 6 yards, a 25-mph crosswind adds 7 yards, the playoffs add 4 yards,
and the difference between the first attempt of the game and the fifth is 7
yards (another surprise).
So far, we have gained valuable information about how much certain factors
affect field goal accuracy. However, what can be learned when the probabilities
are used to evaluate kickers? The first conclusion is that field goal percent-
age does a poor job of rating kickers. The Raiders’ Sebastian Janikowski did
not make a (relatively) high percentage of kicks, but the offensively challenged
Raiders often sent him out to try 55- and 60-yard kicks. Given the difficulty of
his kicks, Janikowski rates as one of the best kickers of his era. That last state-
ment should be read with suspicion. Is the rating persistent with Janikowski
always high on the list, or is the rating essentially random with Janikowski
sometimes floating to the top by luck? Irrespective of Janikowski’s ability, the
sad truth is that the correlation in performance of kickers from one year to
the next was 0.01, essentially zero.
On a relative basis, then, the ranking of kickers has a large component of
randomness. With few kicks being missed, a kick or two off the upright can
drop a kicker well down the rankings. Plus, the kicker has no control over
whether his team needs him frequently from short distances or infrequently
from long distances.
 Big Data and Beyond 211

On the Rebound
What is the best way to evaluate a rebounder in the NBA? Think about
what can be misleading about total rebounds. The team’s pace, the oppo-
nents’ shooting percentage, the positioning of the players, the philosophy of
the coach, and other factors affect the opportunity to gather rebounds. With
player tracking that records the positions of the ball and players at all times,
the possibilities for new rebounding metrics have exploded. The metrics dis-
cussed here come from a paper by Maheswaram, et.al., titled “The Three
Dimensions of Rebounding.”
The overarching concept is, again, Value Added: compare a player’s re-
bounds to the expected number of rebounds for an average player in the same
situation. The researchers divide the rebound into three stages. First is po-
sitioning: given where the players are when the shot is taken, what is the
probability of getting the rebound? Second is hustle: between the time of
the shot and the time of the ball becoming reboundable (reaching the 10-foot
height level), how much has the player improved his or her chances of getting
the rebound? Third is conversion: in what fraction of the player’s rebound
opportunities does the player end up with the rebound?
Imagine watching a replay from an overhead camera. Stop the action when
a shot is taken. Some players are in the lane close to the basket, others are
out at the three-point line. What percentage of time should each player get
the rebound? The first information you need is where the rebound is likely to
occur. Collect enough data and you can estimate the distribution of rebound
locations for shots from different locations on the court. Now, imagine that
a player is standing two feet away from the rim. This is good rebounding
position, right? It depends: is this player the only one around, or are there
other players in tight quarters banging for position? The second information
you need is the amount of area that the player controls.
The second step involves constructing a
Voronoi diagram. Figure 10.2 shows four po-
tential rebounders around the basket. Player D
seems to be in the best rebounding position, but
this depends on where the ball bounces. For now,
our goal is to divide up the court into the regions
that each player controls. The idea is simple: for
each spot on the court, whichever player is clos-
est to the spot controls the spot.
We use basic geometry to divide up the court.
Start with two points, A and B. The dividing
line between A and B is, in fact, a line. If you
mentally sketch in a line that you think splits the FIGURE 10.2: Rebounds
difference between A and B, you will probably
212 Sports Math

get very close to the dividing line. To be precise, though, start by connecting
A and B with a line segment L. The dividing line we want is perpendicular to
L and passes through the midpoint of L.
Figure 10.3 shows the six lines that divide two points at a time, labelled
with the two points being separated. This creates a fairly large mess, but to
get the final picture that we want (the Voronoi diagram) we simply erase a
few of these line segments.
Many of the line segments in Figure 10.3 are
unnecessary. For example, look at the line that
goes just above the letter B in Figure 10.3. This
is the dividing line for points A and C. However,
in this part of the court, player C is not relevant.
This region of the court is well above the line
separating B and C; player B is much closer than
player C, so we do not need to worry whether A
or C is closer. Similarly, look at the portion of the
AD line to the right of the mess of intersections.
Clearly, the choice here is between players B and
C, so the dividing line for players A and D is
irrelevant and can be erased. FIGURE 10.3: Dividers
Figure 10.4 shows the completed Voronoi di-
agram. Example 10.5 below gives you an idea of
how a computer might construct this diagram.
The Voronoi diagram is an integral part of
the computation of the rebounding metrics of
positioning, hustle, and conversion. For the po-
sitioning value, we take the Voronoi diagram and
overlay the likelihood of a rebound coming into
that region. For example, player D has a siz-
able chunk of area in Figure 10.4, but if the
rebound is highly likely to occur in the upper
portion of the diagram, then player D is not in
the best rebounding position. The player’s posi-
tioning value is the probability that the rebound
FIGURE 10.4: Voronoi
(for a shot taken from a particular position on
the court) will occur in the portion of the court that the player controls.
Repeat the calculation after a missed shot has hit the rim. The basis for
the hustle metric is to compute the difference between positioning value and
the new value. This calculation needs to be tweaked, however. Other players
will crowd in to the picture, taking away territory originally controlled by
players A-D. These other players will tend to go to open territory, so the
larger the positioning value is for a player, the more likely that player is to lose
territory (whether that player hustles or not). To be a different “dimension”
of rebounding, we want the correlation of hustle and positioning to be zero, so
the dependence of hustle on positioning needs to be removed. The final hustle
 Big Data and Beyond 213

value, then, is Value Added: the change in rebounding probability compared

to the normal change in rebounding probability for a player with the given
positioning value.
The conversion metric starts by identifying “opportunities.” This is where
a ball is at rebounding height in the player’s Voronoi region. Conversion is
the fraction of opportunities for which the player actually gets the rebound,
above or below the average conversion rate for the given positioning.

Example 10.5 For players A (5,5), B (7,9), and C (6,7), (a) find an equation
of the dividing line between A and B; (b) determine who is closest to the ball
at (10,6).
Solution. (a) The segment between A and B has midpoint (6,7) and slope
9−5
7−5 = 2. The dividing line therefore has slope −1/2 (slopes of perpen-
dicular lines multiply to −1). The line through (6,7) with slope −1/2 has
equation y = 7 − 12 (x − 6). (b) The distance between (10,6) and (5,5)
p √
is p (10 − 5)2 + (6 − 5)2 = 26; The distance between (10,6) and (7,9)
√
is (10 − 7)2 + (6 − 9)2 = 18; The distance between (10,6) and (6,7) is
p √
2 2
(10 − 6) + (6 − 7) = 17. Player C is closest.

It is not the intent of this section to claim that these three measures of
rebounding are the proper way to evaluate rebounding. Instead, the intent
is to give an example of how the new player tracking data can be used to
calculate detailed aspects of the game.

Breaching the Convex Hull

In the 2014-15 regular season, Atlanta’s Kyle Korver averaged 8 field goal
attempts per game, of which nearly two-thirds were catch-and-shoot three
pointers. What is the value of having a player whose primary offensive contri-
bution is to wait to be passed the ball to shoot a three-pointer? Coaches talk
about the importance of “floor spacers” whose excellent long-range shooting
forces the defense to guard them closely. This spreads out the defense and
allows other players to operate more effectively near the basket.
214 Sports Math

In Basketball Analytics: Spatial Tracking, Stephen

Shea develops a way of quantifying the effects of floor
spacers, and details the importance of spreading the
floor in the 2014 NBA Finals. Shea uses a mathemat-
ical construct called a convex hull. Figure 10.5 shows
two defensive alignments. Clearly, the defense on the
bottom is more spread out than the one on top. The
defense on top is packed into the lane, leaving any
potential three point shooters open. The defense on
the bottom is guarding the three point line carefully,
and has left room for drives to the basket and post
moves inside. FIGURE 10.5: Two
The challenge is how to measure the spread of Defenses
the defense. You could measure the “diameter” of
the defense, the greatest distance between any two defenders. However, if one
of the defenders on top moved away from the crowd, you would measure a
larger diameter but you would not want to say that this defense is more spread
out than the defense on the bottom.
Shea uses the area of the convex hull of the de-
fenders to quantify spread. The concept is actually
quite simple. Connect each pair of defenders with a
line segment and then erase interior lines. The re-
sulting polygon is the convex hull. The number of
sides of the polygon is variable. Notice that one of
the defenders in the top defense is on the interior of
the convex hull, whereas the convex hull of the bot-
tom defense is a pentagon. The area of the convex
hull of the packed defense is clearly much smaller
than that of the spread-out defense.
Shea computed areas of convex hulls of the San
Antonio defense against Miami in the 2014 NBA Fi- FIGURE 10.6: Two
nals. The areas quantify how the San Antonio de- Convex Hulls
fense clogged the lanes in the first three games and shut down Miami’s of-
fense. However, in game four increased playing time for Miami floor spacers
Ray Allen, Shane Battier, and others forced San Antonio to spread its defense.
Miami’s offensive efficiency improved, and the Heat came from behind to win
the championship.

Example 10.6 Compute the area of the convex hull of the points A (1,3), B
(4,2), C (3,6), and D (2,4).
Solution The convex hull is the triangle ABC (sketch a quick graph to con-
vince yourself that this is true). We compute the area of the triangle ABC
using Heron’s formula, which only depends
√ on the
√ lengths of the sides. The
distance between √ A and B is a√ = 32 + 12 = 10, the distance between
B and C is b = 12 + 42 = 17, the distance between A and C is c =
 Big Data and Beyond 215

√ √ p
22 + 32 = 13. Heron’s formula gives the area as s(s − a)(s − b)(s − c)
where s = (a + b + c)/2. In this case, the area equals 5.5.

Calculus Box: A Goal-Scoring Model

The various Value Added metrics that we have seen depend on accurate
estimates of what is expected to happen. Expected values examined in this
chapter have come from massive data sets giving reasonable empirical esti-
mates. Other expectation calculations may come from a more sophisticated
use of statistics than has been presented in this book. The next example is of
that type.
In soccer and hockey, scores are low and goals can only be scored one at
a time. These are indicators that a statistical distribution called the Poisson
λn −λ
distribution may apply. The distribution function is e , which gives the
n!
probability that n goals will be scored in one unit of time, if an average of λ
goals are scored in that (arbitrary) unit of time.

Example 10.7 A soccer league has an average of 2.7 goals scored per game.
Use the Poisson distribution to estimate the probability that (a) 4 goals are
scored in a game; (b) 0 goals are scored in a game; (c) 1 goal is scored in the
second half of a game that is tied 1-1 at half.
Solution The average for a game is λ = 2.7. (a) The probability of n = 4 goals
2.74 −2.7
scored is e ≈ 0.149, about a 15% chance. (b) The probability of n = 0
4!
0
2.7 −2.7
goals scored is e = e−2.7 ≈ 0.067, about a 7% chance. (c) The score
0!
is irrelevant. By the Poisson model, the probability of getting n = 1 goal in a
1.351 −1.35
period of time in which the average is λ = 1.35 goals is e ≈ 0.350,
1!
about a 35% chance.

Part (c) of Example 10.7 points out one of the important hypotheses under-
lying the Poisson distribution: the timing of goals is independent. Whether
the score is 0-0 or 1-1 or 4-0, the probability of n more goals being scored
is governed by the Poisson formula. You may think that the independence
assumption is not valid. However, in The Numbers Game Chris Anderson de-
scribes research debunking the theory that teams are scored upon more often
immediately after scoring a goal than at other times. In fact, the goals seem
to come independently during the major portion of the game.
216 Sports Math

Figure 10.7 compares the actual number of Goals Actual Pred.

goals scored in the 380 EPL games in 2014-15 to 0 31 29
predictions of goals scored using a Poisson model 1 77 75
with an average of λ = 2.57 goals per game (data 2 88 96
from Soccer STATS). The match is quite good. 3 85 82
The most significant mismatch is that fewer two- 4 56 53
goal games were played than predicted. The sta- 5 27 27
tistical significance of this mismatch is explored 6 10 12
in Example 10.8. Further details about the mis- 7 or more 6 6
match can be found in Figure 10.8. In the dis-
cussion of independence above, you may have FIGURE 10.7: EPL
thought that all bets are off near the ends of
games. A team down 1-0 may throw caution to the wind trying to get the
equalizing goal. While this might produce more 1-1 games, it could backfire
and produce more 2-0 games. Similarly, in a 1-1 game will both teams play
conservatively and take the point or will they push forward to get a winning
goal? The results are given next.
Figure 10.8 breaks down the scoring Home Away Actual Pred.
into home and away goals. The 2014- 1 0 40 43
15 average for home goals is λ1 = 1.47 1 1 37 47
and the average for away goals is λ2 = 2 0 37 32
1.09. The eight most common scores are 0 1 37 32
shown. The match is good, but not as 2 1 35 35
precise as in Figure 10.7. Given that 0 0 31 29
there are many more possible outcomes 1 2 26 26
with smaller sample size, this is not sur- 2 2 22 19
prising. In terms of two-goal games, note
the large shortfall in 1-1 games compared FIGURE 10.8: EPL
to the model prediction. This is the main
contributor to the deficit in two-goal games seen in Figure 10.7. The numbers
for 0-2 games are not shown, but there were 14 such games compared to a
prediction of 17 games.

Example 10.8 Compute the probability, for a Poisson process with λ = 2.57,
that 2 goals would be scored in 88 or fewer games out of 380.
Solution Under the assumptions, the probability of scoring two goals is
2.572 −2.57
e ≈ 0.2528. Treating games as Bernoulli trials, the probability of
2!
getting 2 goals in exactly 88 out of 380 games is 380

88 292
88 (.2528) (.7472) ≈
0.03
and the probability of 88 times or fewer is the sum from 0 to 88 of
380 n 380−n
n (.2528) (.7472) , which is approximately 0.187. An occurrence of 88
or fewer 2-goal games would occur by chance about 19% of the time under a
Poisson model, so the largest deviation from predicted values in Figure 10.7
is not statistically significant.
 Big Data and Beyond 217

Showing Hot and Cold

The collection and analysis of massive amounts of data is of little use un-
less interesting results are effectively communicated. Sophisticated computer
graphics play an important role in clarifying patterns. Heat maps have become
a common way of displaying data. They lose some impact in the black and
white of this book, but some comments may be useful.
A heat map displays data, generally two-dimensional, using colors. The
choice of color palette is entirely arbitrary, although when creating heat maps
you should keep in mind that reds and oranges are normally associated with
“hot.” Here is a simple example.

Example 10.9 When the strike zone is divided into nine regions, a batter
has
 the given batting averages in the regions. Display this as a heat map.
.310 .350 .290
.280 .320 .270
.260 .260 .220
Solution We convert each number to a color. In color, you could range from
.220 in blue to .350 in red. For black and white, the range could be light gray
for .220 and black for .350. This is shown in Figure 10.9a.

FIGURE 10.9: Heat Maps for Example 10.9

For a web page, the point of a heat map may be to provide a colorful
distraction, but if the point is to convey information labels are useful. If you
came across Figure 10.9a out of context, could you tell that the darker regions
are the batter’s hot zones? In Figure 10.9b, the averages are included to remove
ambiguity from the graphic. Notice also that the change from black to white
font draws attention to the regions with averages over .300.
218 Sports Math

A classic illustration of the power of graph-

ics involves the display of pitch locations. The
New York Times video “How Mariano Rivera
Dominates Hitters” not only gives excellent in-
formation on Rivera’s abilities, it also gives an
invaluable lesson in communicating with graph-
ics. The next two figures make the same point. FIGURE 10.10: Pitches
Figure 10.10 displays the locations of three hun-
dred simulated pitches. We display each pitch as a baseball. The “clever” use
of the baseball image may be eye-catching, but little can be learned from the
mess of balls.
For Figure 10.11, we tone down the cute-
ness and display each pitch as a see-through
(opacity 0.2) disk. With the low opacity, you
can see where most of the pitches end up. We
can now learn that this pitcher threw a high
percentage of pitches on the two corners of the
plate. The pitcher’s control and effectiveness are
demonstrated by the graphic’s extra dimension FIGURE 10.11: Pitches
in which we can see the balls stacking up on the corners.

RIP to the RPI

The analytics movement in sports will have several indirect effects as quan-
titative thinking spreads throughout the system. An example of this involves
Roanoke College, a Division III college with a long history of excellent soc-
cer teams. Despite its history and the strong performance of its conference in
the NCAA tournament, at-large bids to the tournament became increasingly
scarce for members of the conference.
At-large bids are determined by a selection committee, which gives large
consideration to the RPI (Rating Percentage Index). The RPI for a team
is 25% of the team’s winning percentage plus 50% of its opponents’ winning
percentage plus 25% of opponents’ opponents’ winning percentage. Roanoke
coach Ryan Pflugrad thought to analyze the effect of Roanoke’s large confer-
ence size (12 teams) on RPI.

Example 10.10 Compute the RPI for teams A and B: (a) A’s record is
14-3, its opponents’ record is 157-132, and its opponents’ opponents’ record
is 1200-800; (b) B’s record is 14-3, its opponents’ record is 164-125, and its
opponents’ opponents’ record is 1200-800.
Solution (a) Winning proportions are .8235 (14/17), .5433 (157/289), and
.6000 (1200/2000) so A’s RPI is .25(.8235) + .5(.5433) + .25(.6) = .6215.
 Big Data and Beyond 219

(b) Winning proportions are .8235 (14/17), .5675 (164/289), and .6000
(1200/2000) so B’s RPI is .25(.8235) + .5(.5675) + .25(.6) = .6336. Despite
the identical 14-3 records and 60% opponents’ opponents’ records, B has a
significantly higher RPI because its opponents won 7 more games than A’s
opponents, indicating that B played a tougher schedule.

Most likely, the outcome of Example 10.10 does not strike you as unrea-
sonable. RPI rewards the team that played the tougher schedule. The lessons
from Chapter 5 should make you a little suspicious of a system that only
uses wins and losses, but the two components of RPI that attempt to quan-
tify strength of schedule might ease that suspicion. Put into proper context,
however, Example 10.10 actually reveals a major flaw in the RPI.
Consider two teams of equal ability and equal strength of schedule, except
that team A plays in a 12-team conference and team B plays in an 8-team
conference. To keep the comparison simple, suppose that both teams go 14-3,
and that their opponents’ opponents’ records are identical 1200-800 marks.
So, we restrict the difference in RPIs to the opponents’ records. Team A plays
11 conference games and 6 out-of-conference games, while team B plays 7
conference games and 10 out-of-conference games. Suppose that A goes 10-1
in conference and 4-2 out of conference, and B goes 6-1 in conference and 8-2
out of conference.
Table 10.1 summarizes the records of A’s and B’s opponents. For both, the
out-of-conference opponents (OC) win about 60% of the games: 61-41 for A’s
6 opponents, 100-70 for B’s 10 opponents (100-70 actually rounds to 59%).
For both, the conference opponents win 60% of their out-of-conference games
(C-OC): A’s 11 opponents go 40-26 (6 games each), B’s 7 opponents go 42-28
(10 games each). Finally, the conference schedules are equal in the sense that
every team in each conference plays all other teams exactly once. In conference,
then, A’s 11 opponents play 10 games against non-A teams; in each game, one
team (not A) wins and one team (not A) loses so the records sum to 55-55.
Since A went 10-1 in conference, the other teams went 1-10 against A, making
the total conference record 56-65 (shown in the C-C column). Similarly, B’s
conference opponents go 21-21 plus 1-6 for a total of 22-27.
TABLE 10.1: Records of Opponents
OC C-OC C-C Total
Team A 61-41 40-26 56-65 157-132
Team B 100-70 42-28 22-27 164-125

You have undoubtedly noticed that teams A and B in this scenario are
the teams A and B in Example 10.10. The apparently tougher schedule of
team B is simply an artifact of being in a smaller conference. We see that a
team’s excellent conference record harms the records of its opponents. If team
A plays more conference games than B, A’s strength of schedule is harmed
220 Sports Math

more than B’s. Example 10.10 shows that the difference between a 12-team
conference and an 8-team conference is significant.
This calculation oversimplifies the effects of playing more or less conference
games, but there is a clear lesson that the RPI penalizes teams in larger con-
ferences. The punchline: Roanoke’s conference, presented with the analytics
argument, voted to split (for soccer) into two six-team divisions.

Blackbox Analytics
While it is true that computers merely follow instructions, when the in-
structions include ways to modify the instructions and evolve, amazing and
unpredictable results can occur. Most of us have felt the wrath of a computer
virus, a harmful example of an “autonomous” program, and you have prob-
ably heard of machine learning. We next take a quick look at neural nets, a
popular type of machine learning.
Suppose that we want to use Four Factors to predict which team wins a bas-
ketball game. Recall that the four factors are shooting, turnovers, rebounds,
and free throws. (As presented in Chapter 7, Four Factors was retrodictive;
that is, we used the results of the game itself to retroactively evaluate who
should have won the game. We could also use Four Factors to predict the
outcome of a game, plugging in season averages for each team.) The ques-
tion is how to combine the four inputs into a single number that predicts the
outcome of the game.
A neural network typically consists of three layers: inputs, hidden, and
outputs. We can think of having four inputs (the difference in the teams’
shooting, and so on). For Example 7.7, the four inputs would be 0.032 (San
Antonio’s net shooting), −0.018 (turnovers), −0.92 (rebounds), and −0.116
(free throws). These four inputs are fed into the hidden layer.

FIGURE 10.12: Neural Network

The general neural network in Figure 10.12 shows two columns of dots or
nodes (the number of nodes is arbitrary) in the hidden layer. The idea is that
 Big Data and Beyond 221

each input could be entered into each node, and each node could compute a
different function of the inputs. Outputs from the first set of nodes could be
fed into a second column of nodes, with each of these nodes creating its own
combination of its inputs. The final column of nodes could then be combined
into the final output. The hidden layer of the network can be very complicated,
such that it can be quite difficult to trace the path of a single input through all
of the nodes in the hidden layer and determine its effect on the final output.
We have choices to make for the structure of the hidden layer. The simplest
is to compute the weighted sum of the inputs and output “San Antonio wins” if
the sum is positive, or “Los Angeles wins” if the sum is negative. For example,
if the weights are 25, −10, 1, and 0.5, we compute 25(0.032) − 10(−0.018) +
(−0.92) + 0.5(−0.116) = 0.002. The positive value indicates that we predict a
San Antonio win (which is correct; the Spurs won, 111-107). This is equivalent
to a linear regression. We could use the small size of the output to say that
the game was very close and output a probability. Using a sigmoid function,
1
we could output 1+e−10∗0.002 = 0.505 and say that San Antonio had a 50.5%
chance of winning the game.
Neural networks are modeled on the action of neurons which fire when a
threshold is reached. With this in mind, we could compare each input to a
threshold and send an output of 1 or −1 if the difference is significant enough.
For example, a shooting difference of 0.032 might be highly indicative of a
winning team, a rebounding difference of −0.92 indicative of a loser, and
the turnover and free throw values too close to call. The output would be
1 + 0 − 1 + 0 = 0, and we would not predict a winner. Or, each input could
be plugged into a sigmoid function to produce a probability (e.g., a team that
has a net shooting value of 0.032 wins 74% of the time) and the probabilities
averaged.
Whichever choices we make for the structure of the neural network, there
are constants to be determined (e.g., the weights 25, −10, 1, and 0.5) in the
hidden part of the network. The magic of a neural network is that the network
itself determines these values by training on large numbers of examples (e.g.,
for the current values of the constants, predict the San Antonio-Los Angeles
game and 100 other games, and see how many predictions are right). The
constants are tweaked to improve the predictions as much as possible. Practi-
tioners generally use multiple training sets to try to avoid the neural network
adjusting itself to oddities of the training set. A famous example involves a
neural network that was supposed to learn to distinguish satellite pictures
of tanks from pictures of civilian vehicles, but instead learned to distinguish
sunny from overcast days because all of the pictures of tanks in the training
set were taken on overcast days.
The programmer of a neural network does not have to know much about
how the inputs relate to the output; the network figures that out. So, a belief
that the Four Factors can be used to predict winners is enough to run a
model. Depending on how many hidden layers are used and what type of
thresholding is employed, the neural network may find patterns that would
222 Sports Math

have eluded human observers. The patterns may be complicated enough that
it is difficult to explain what is going on, even after the fact. This can be a
distinct disadvantage of neural networks: if the analyst cannot explain how the
network is getting its results, the player or manager may dismiss the results
out of hand.
Neural networks have been applied to the complex problem of predicting
the next baseball pitch. As such models increase in accuracy, you can imagine
an arms race of sorts. Would it be unethical or illegal to transmit such predic-
tions to a batter? Would the catcher and pitcher have a similar network telling
them what the batter is likely to be looking for? The game theory discussion in
Chapter 9 comes into play here, in that optimal strategies are supposed to be
implemented randomly. This would foil any such pitch prediction algorithms.
The future may be highly mathematical!
Neural networks and other machine learning techniques are powerful and
flexible tools for discovering relationships in complex networks.

PeeWee Analytics
Figure 10.13 shows the trajectory and some
data from a recent golf swing of mine at the driv-
ing range. My backswing is more extended than my
downswing. The “ratio” of 4.6 measures tempo and
equals the ratio of the time spent in the backswing
to the downswing. Other data, recorded but not
shown here, include a ball launch angle of 12.3◦ ,
a closed clubface angle of 3.9◦ , an attack angle
of 0.9◦ , and a plane angle of 44◦ . Some of this
data may create paralysis by analysis, but having
a record of what my swing looked like when I was FIGURE 10.13: Drive
hitting the ball well could prove invaluable.
In addition, any golfer can now buy products that record the details of
every shot on the course. In the clubhouse, traces of where every shot went
(overlayed on a picture of the hole being played) can be printed out and
statistical analyses of Strokes Gained putting, driving, and so on displayed.
All of the data collection and analysis tools discussed in this chapter, plus
many more, will be available soon to coaches in 10-year-old recreation leagues
of all sports. Whether this is a good or a bad development is debatable: I would
have loved this when I was 10, but it probably would have done more for my
development as a mathematician than for my development as an athlete.
 Big Data and Beyond 223

Wearable Tech
Imagine a quarterback preparing for a game with a game simulator, some-
what like Madden Football viewed on your sunglasses. He sees the opponents’
blitzes and zone defenses in realistic three-dimensional graphics that fully
prepare him for the upcoming game. The quarterback coach reviews the quar-
terback’s eye movements during the simulation and reminds him (again) that
he needs to look off the linebackers on passes over the middle. The sensors
in the quarterback’s shoes record a left/right force imbalance that indicates
that the quarterback’s left knee is still sore. For the game, a patch will be
applied to the knee that monitors the functioning of the knee and delivers a
pain suppressant as needed.
The backup quarterback’s simulator session did not go as well, as he fielded
several phone calls on his simulation glasses. His jersey recorded an unusually
rapid buildup of lactic acid, indicating that he is still out of shape. His hy-
dration levels were also abnormal, possibly due to aftereffects of the party the
night before. His eye movements and response times showed lingering effects
of the concussion he had suffered in a previous game.
Does this sound like fiction? Perhaps by the time you read this, it will be
commonplace. The scenarios required little imagination, as all of this technol-
ogy exists in 2015.
The most reliable prediction for the future in sports is that technology will
revolutionize every aspect of the athletes’ and spectators’ experience, in ways
that we cannot imagine today.

Exercises
10.1 Discuss how data such as catch-and-shoot percentage, pull-up shooting per-
centage, drive percentage, and so on could affect defensive strategy in basketball.
10.2 Discuss how data such as miles run, speed, and acceleration could affect
substitution strategy in soccer.
10.3 Discuss how charts showing where shots are hit from and to could affect
service and return strategies in tennis.
10.4 Discuss how wearable tech in shoes, shirts, and so on, in practice could
be used by coaches in making decisions about demoting starters or changing
substitution patterns.
10.5 (a) Repeat Example 10.2 with the approach shot finishing 60 feet from
the hole. Explain why the overall Strokes Gained is the same even though the
approach shot and putt values have changed. (b) Compute Strokes Gained for
each shot on a 380-yard par 4. A 300-yard drive in the fairway is followed by an
80-yard approach shot 20 feet from the hole, then two putts.
224 Sports Math

10.6 Estimate the caught stealing rate for a (a) pop time of 1.98 s; (b) delivery
time of 1.3 s; (c) running time of 3.4 s. (d) Use your answers to (a)-(c) and
Example 10.3 to estimate the effect of 0.1 s on the caught stealing rate.
10.7 (a) Use h(x) from Example 10.3 to estimate the running time needed to
have a steal rate of 80%. (b) For a pop time of 1.8 s and delivery time of 1.2 s,
estimate the caught stealing rate. (c) For a pitcher/catcher duo with the values
of part b, is your estimate from part a too high or too low?
10.8 At the Fielding Bible’s rate of 0.12 runs per strike, how many extra strikes
would a catcher need to frame to save his team (a) 30 runs? (b) 300 runs? (c)
Do you think framing can save 300 runs in a season?
10.9 (a) Use the distance model from Example 10.4 to determine the distance at
which NFL kickers would make 50% of their field goals. Does it seem realistic?
(b) Repeat for 25%. (c) Repeat for 10%.
10.10 An NFL kicker who misses 3 or 4 field goals in a row is likely to be dis-
missed. Describe how this could cause the low persistence of 0.01 in Pasteur’s
field goal kicker metric.
10.11 Find an equation of the dividing line for each pair of points. (a) (4, 1) and
(4, −3); (b) (4, 1) and (6, −1); (c) (4, 1) and (6, 5).
10.12 Sketch the Voronoi diagram for players at points A(4, 1), B(4, −3), and
C(6, −1) (use exercise 10.11).
10.13 Determine which player is closest to the ball at point P, player A(4, 0),
B(6, −3), or C(5, −2) (a) P is (8, 1) (b) P is (2, −5).
10.14 Repeat Example 10.6 moving C to (3,7). Is the defense more spread out?
10.15 Compute the area of the convex hull of the points A (1,0), B (4,2), C (3,1),
and D (5,1).
 
.230 .270 .220
10.16 Draw a heat map for the batting averages shown: .280 .340 .310
.280 .330 .360
10.17 Find heat map-like graphics in at least three different sports (links pro-
vided at the book’s website). Critique the presentation of each. Is it clear which
colors represent good play? Is it clear which regions have a large sample size for
a “hot” or “cold” designation to be significant? Is it giving you information that
you want?
10.18 Compute the RPIs of teams with the given record, opponents’ record, and
opponents’ opponents’ record. (a) 10-1, 58-63, 312-380 (b) 8-4, 80-64, 510-422
10.19 Keeping opponents’ records the same, in Example 10.10 what would B’s
record have to be to make its RPI worse than A’s?
10.20 For Table 10.1, we adjusted C-C to account for the effect of A’s and B’s
records. We did not adjust OC or C-OC, however. (a) Explain why C-OC does
not need to be adjusted. Speculate on whether C-OC would be better or worse if
each team played more games. (b) To adjust OC, start with a 60% mark for each
team’s opponents, 57-38 for A and 96-64 for B. Add in the 2-4 mark against A
and 2-8 against B and recompute the table. Does B’s advantage over decrease?
10.21 T The various Value Added statistics depend on large data sets and
calculations that the average fan could not do by hand. Discuss whether this
reduces the reliability of the stat, or the charm of the stat, or the likelihood of
the stat becoming accepted by fans.
 Big Data and Beyond 225

10.22 T Baseball’s WAR statistics are value added “above replacement” play-
ers, while golf’s Strokes Gained is strokes better or worse than average. Discuss
which standard is more appropriate, the average value or the value of a replace-
ment (potentially minor league) player.
10.23 T Suppose that you hear a football analytics expert say that successful
long passes are more dependent on the receiver than on the passer, while suc-
cessful short passes are more dependent on the passer. Explain in football terms
why this might be true. Discuss the evidence that might be collected to back up
such a statement.
10.24 T ESPN’s Total QBR attempts to rate quarterbacks on all aspects of
their play, including running and passing. Discuss whether the rating should
depend on “leverage” (the importance of the play as measured by changes in
win probabilities).
10.25 T Discuss ways in which knowledge of delivery times and pop times could
be used to decide whether to try to steal a base or not.
10.26 T Find a video showing catchers framing pitches (e.g., http:// grant-
land.com /features/studying-art-pitch-framing-catchers-such-francisco-cervelli-
chris- stewart-jose-molina-others/ accessed 9/11/2015) and describe what the
catcher is doing.
10.27 T It was reported that the field goal kicker evaluation metric has a per-
sistence of 0.01. Discuss the meaning of this: field goal kicking at the NFL level
is not consistent, the metric is flawed, or something else.
10.28 T If RPI penalizes teams in larger conferences, should it be used to help
rank teams for a tournament?
10.29 T In basketball, is it easier to get an offensive rebound on a long or short
shot? Basketball on Paper reports that NBA teams get offensive rebounds on
33% of 2-point shots and 31% of 3-point shots. Discuss the extent to which this
answers the question posed.
10.30 C (a) Show that the function p in Example 10.4 is decreasing and explain
why this is important. Find limits of p as the distance goes to (b) 0; (c) infinity.
(d) Which is these values is not realistic?
10.31 C Suppose that the distribution of rebounds for a shot is inversely pro-
portional to the distance from the rim for 1 < d < 15. Computepa double integral
to find the value of k in the distribution function f (x, y) = k/ x2 + y 2 (assum-
ing that the basket is at the origin and the foul line is at x = 15). If player A is
at (2, −1) and player B is at (4, 1), find the positioning values for each player.
10.32 C Compute the area of the convex hull of A (0,0), B (1,2), C (4,2), D
(5,0) in two ways. (a) Show that the convex hull is a trapezoid and compute its
area. (b) Find the areas of each of the four triangles formed and combine them
in the appropriate way.
10.33 C Repeat Example 10.7 for hockey’s average of 5.32 goals per game.
10.34 C NHL hockey games average 5.32 goals per game, so it has been said
that the NHL is a “3-2 league.” (a) Using the home team average of 2.78 goals per
game and the road team average of 2.54 goals per game, compute the probability
that a game ends 3-2. (b) Is 3-2 the most likely score?
226 Sports Math

10.35 P Construct a neural net model for predicting basketball games. (See,
for example, Loeffelholz et.al. in JQAS 2009.)
10.36 P Research the methodology of genetic algorithms and construct a ma-
chine learning algorithm for predicting basketball games.
10.37 P Find data on the number of baseball pitches of different types thrown
by pitchers, and compute the pitch-type entropy for each pitcher. What can
a batter expect from a high-entropy pitcher? Investigate whether high-entropy
pitchers are as a group better, worse, or the same as low-entropy pitchers.
10.38 P The Hardball Times 2014 Baseball Annual reports that pitcher’s veloc-
ity in a game peaks in the first 20 pitches, then gradually declines for subsequent
pitches. Investigate this claim.

Further Reading
A list of data-rich websites will be maintained at this book’s web site. In
2015, the official league sites all have good information, and the reference.com
sites (e.g., baseball-reference.com) are excellent.
Golf By the Numbers presents a development of Strokes Gained.
ESPN’s explanation of its Total QBR system is at http://espn.go.com/ nfl/
story/ /id/6833215/explaining-statistics-total-quarterback-rating accessed 9-
14-15.
Annual guides include The Fielding Bible from Baseball Info Systems, the
Football Outsiders Almanac, Hardball Times’ Baseball Annual, and guides
from the Prospectus family (Baseball, Basketball and Hockey).
Journals include the online Journal of Quantitative Analysis of Sports and
European Journal of Sport Science. The MIT Sloan Sports Analytics Confer-
ence posts numerous research articles and videos.
Severini’s Analytic Methods in Sports gives an excellent overview of statis-
tical methods applied to sports. Other excellent books about sports analytics
include Baseball Between the Numbers by Baseball Prospectus, The Sabermet-
ric Revolution by Baumer and Zimbalist, Mathletics by Winston, Basketball
on Paper by Oliver, Basketball Analytics Spatial Tracking by Shea, Analyz-
ing Wimbledon by Klaasen and Magnus, Stumbling on Wins by Berri and
Schmidt, and The Book by Tango and Lichtman.
“Who Is Responsible for a Called Strike?” Joe Rosales, Scott Spratt, SSAC
2015 proceedings accessed 9-10-15
An article on using data analysis to predict baseball pitches:
http://www.sporttechie.com /2014/03/03/applying-data-science-to-predict-
mlb-pitching-patterns/ accessed 8-24-15
Answers and Selected Solutions

Chapter 1: Projectile Motion

1.1 We have c = −9.8 , v0 = 0 m/s and p0 = 100 m. Then v(t) = −9.8t
m/s and p(t) = 100 − 4.9t2 . At t = 1, v = −9.8 m/s and p = 95.1 m. At
t =p2, v = −19.6 m/s and p = 80.4 m. At impact, p = 0 m. This occurs at
t = 100/4.9 ≈ 4.52 s, at which time v ≈ −44.3 m/s.
√ √
1.2 (a) −32 2 ft/s (b) −64 2 ft/s (c) 2

1.3 96.1 m, 80.4 m, 55.9 m, 21.6 m; −4.9 m, −14.7 m, −24.5 m, −34.3 m;

9dp1 = −44.1 m, 11dp1 = −53.9 m; 3, 5, 7, 9, ...

1.4 (a) t = 1: p = 6.1 m, v = −8.8 m/s; t = 2: p = −7.6 m, v = −18.6 m/s

(b) h = 10.051 m, t = 1.534 s, v = −14.036 m/s

1.5 (a) 14.51 ft/s (b) 0.907 s (c) 4 ft

1.6 (a) 80 ft/s (b) 100 ft

1.7 (a) good (if entry angle ok) (b) not a swish, goes off backboard

1.8 29.9 ft/s to 30.2 ft/s

1.9 24.941 and 24.938; minimum force

1.10 (a) hits ground before base (b)3.65◦ , 5.1 ft

1.11 (a) no, too long (b) 7.7◦ to 9.0◦ (c) 8.0◦ to 9.2◦ , small decrease

1.12 y = 5.1, no

1.13 (a) yes (b) no

1.14 108.2 mph, about 14% more

1.15 k ≈ 0.00304, 112.5 mph

1.16 k ≈ 0.212, 59.2 km/hr

227
228 Answers and Selected Solutions

1.17 (a) spin left, Magnus down, ball drops (b) spin down, Magnus right, ball
moves right (c) spin up and right, Magnus up and left, ball goes higher and
moves left (d) spin down and left, Magnus up and left, ball hooks left (e) spin
left, Magnus up, ball goes higher

1.18 5.29 m/s, 10.1 m/s, 11.11 m/s, 11.63 m/s, 12.05 m/s, 12.19 m/s, 12.35
m/s, 12.19 m/s, 12.05 m/s, 12.05 m/s; 10 m/s, 11 m/s, 11.5 m/s, 12 m/s, 12.2
m/s, 12.3 m/s, 12.3 m/s, 12.2 m/s, 12.1 m/s, 12.1 m/s; 10 meters; 10 m/s2 ,
1 m/s2 , .5 m/s2 , .5 m/s2 , .2 m/s2 , .1 m/s2 , 0 m/s2 , −.1 m/s2 , −.1 m/s2 , 0
m/s2 ,

1.19 14.28 m/s, 13.49 m/s; fatigue; 13.89 m/s; 14.70 m/s, drafting; 7.01 m/s,
6.09 m/s; less friction, air drag; fatigue

1.20 (a) 10.38 m/s, 10.86 m/s, running start (b) 9.095 m/s, 9.091 m/s, slower
runners in relay

1.21 7.01 m/s, 9.09 m/s, fatigue

1.22 men 19.69d−.13 women 18d−.133

1.23 157.5 mph

1.24 about 33% less

30
1.25 π ω rpm, 15 rev

1.26 Small area to produce spin, more pressure from fingers to roll ball

1.27 Since x = ct, can replace label x on x-axis with x/c on t-axis

1.28 Better testing for steroids, blood doping

1.29 (a) back, zero, forward (b) Magnus force pushes horizontally (c) Magnus
force pushes ball forward at end

1.30 Longer contact with bat gives more spin gives more distance

1.31 Topspin up, Magnus down; curveball up right, Magnus down left

1.32 Ball rolls up instep so spin up right, Magnus up left; opposite

1.33 Swing plane tilts to right, spin up right, Magnus force up left

1.34 Less surface area on ball makes less drag.

1.35 (a) Changes direction too rapidly (b) Less drag but less Magnus force

1.36 Friction between feet and track, air drag

 Answers and Selected Solutions 229

1.37 reduced air drag

1.38 more contact means more spin, more Magnus force, shots curve more

1.39 horizontal air drag is not balanced by gravity

1.40 Swimmers get speed by pulling their hands through the water. Pulling
against a thick medium creates more force than against a thin medium.

1.41 The reactive force of the floor from the heavy soccer ball is passed on to
the light racquetball, which bounces high.

1.42 For a putt moving up and right, the gravity vector pointed straight down
is to the right of the velocity, accelerating the ball farther to the right. Moving
downhill, gravity points to the left of the velocity vector, helping to straighten
out the putt.

1.43 v = c − ce−t + v0 , p = ce−t + (v0 + c)t + p0 − c

1.44 slope −0.949, angle −43.5◦ , x = 24.558 ft, 0.447 > 0.346

1.45 max at about ω = 1.2 rad/s, limit is 0

1.47 126.7 ft/s vs 127.2 ft/s

1.49 0.389 s is less than 0.395 s; 93 mph pitch is effectively faster.

Chapter 2: Rotational Motion

2.1 (a) We have θ(0) = θ0 rad and θ(T ) = θ0 + 2π rad. Assuming that ω is
constant, θ(t) = at + b for constants a and b to be determined. Then θ(0) = b
and since θ(0) = θ0 , we have b = θ0 . Then θ(T ) = aT + θ0 = θ0 + 2π and so
aT = 2π and a = 2π 2π
T . This gives us θ(t) = T t + θ0 . (b) We have θ(0) = θ0
rad and θ(T ) = θ0 − 2π rad. The steps in part (a) lead to θ(t) = − 2π
T t + θ0 .

2.2 60n rpm, 2πn rad/s

√ √
2.3 B: 20% more, C: 9.5% more ( 2.4 ≈ 1.095 2)
p √ p √
2.4 (a) t = π/2 s, ω = 8π rad/s (b) t = 2π/5 s, ω = 10π rad/s; 25%
increase in α but 12% increase in ω

2.5 9.71 rad/s and 11.89 rad/s (22% more); not realistic

2.6 (a) t = 1.944 s, 15.7 rad/s (b) 5.2 rad/s

230 Answers and Selected Solutions

2.7 1.8 rad/s, less time to catch balance

2.8 436 lb
2.9 120 < 144 so defender has larger linear momentum; −1.6 ft/s
2.10 More than twice as fast; keep moving legs or push with arms
p
2.11 27/2 m
2.12 f (14)/f (13) ≈ 1.239, increase from 12 to 13 is more higher percentage
increase than 13 to 14.
2.13 144ρπ and 169ρπ (17% more); decrease density
2.14 3/α for pike, 4.61/α (50% longer) for layout
2.15 15/3.5 ≈ 4.3 times faster
2.16 c ≈ 2.96
2.17 more weight near rotation point
2.18 longer rotation radius gives more clubhead/racket speed
2.19 cannot keep pushing/pulling legs and arms; simple approximation
2.20 Spiral has smaller MOI, making it easier to throw with high spin rate.
2.21 Harder to get high spin rate, but ball retains spin rate longer.
2.22 (a) PWS helps for serve, not for ground stroke; (b) PWS helps for ground
stroke, not serve
2.23 chest is closer to center of mass than ankle so less torque
2.24 changing distance has more impact
2.25 increases MOI; from above, far left and right
2.26 reduces the MOI for flips
2.27 body rotates in the opposite direction
2.28 smaller MOI for a rotation, so easier
2.29 want smaller MOI when large rotation rates are good, e.g. swinging bat
p
2.30 (a) cπ[(121/192)L3 − (121/192)243 ] (b) at t = 2π/α (e) for 30 < L <
35, the graph of 48 − .34x is higher (f) constant α, shape of bat
1
2.31 4 cπ[(R + 2)4 − R4 ]
2.32 approximately 8880c
 Answers and Selected Solutions 231

Chapter 3: Sports Illusions

3.1 In ft/s, s = 75 ∗ 528/360 = 110 so θ0 = 2∗110
4+x2 which has a maximum with
x = 0 of 55 rad/s, well above the 3 rad/s limit.
√
3.2 (a) 6 ft/s (about 4 mph) (b) 84 ≈ 9 ft
30∗154
3.3 (a) 77 rad/s (b) 9.9 ft (c) max rate of 302 ≈ 5 rad/s, so no
300∗264
3.4 (a) 3002 < 1 rad/s (b) 88 ft

3.5 (a) 0.41 s (b) 0.20 s (c) 0.31 s (d) 0.23 s (e) 0.39 s

3.6 8.3%; 2.0%; 364 days

3.7 (a) 18.6◦ vs 15.8◦ , 11.9◦ vs 12.0◦ (better!) (b) 17.1◦ vs 13.9◦ , 10.0◦ vs 9.5◦

3.8 (a) 1 yard, 9 inches (b) 3 yards, 9 inches

3.9 7.34 m, larger than record 6.16 m

3.10 position change is 2 m > 1.5 m

1 1
3.11 if D is at (0,0) and AR at (40,2) then (a) 0 < y < 20 x (b) 0 > y > 20 x

1 1
3.12 (a) 0 < y < 40 x (b) 0 > y > 40 x

3.13 at 10 m/s will move .5 m, could start on by .25 m, finish off by .25 m

3.14 1968 not special; Olympics may better reflect typical performance

3.15 The ball suddenly appears in an unexpected location.

3.16 no; he could have jumped gaze ahead to where ball meets bat

3.17 an unusual angle, few visual clues.

3.18 little experience deciphering pitch movement

p
3.19 t = 2d/g is reaction time if stick drops d units

3.20 She anticipated the green light and has great timing.

3.21 pressure sensors on starting blocks; .1 s is less than best reaction time

3.22 nature: helpful physical traits; nurture: better training

3.23 long arms to grab ball, reach higher

232 Answers and Selected Solutions

3.24 moving heavy, light weights involve different muscles

3.25 transition from backswing to forward swing bends elbow, giving poten-
tial energy that increases racket speed

3.26 (b) because both players will be seen to have moved farther

3.28 flash-lag says the runner would appear farther ahead and therefore safe

3.29 the best jumpers had grown up using the flop

3.30 θ0 (t) = 1
1+(x/2)2 ∗ (x/2)0 = 2s
4+x2

Ls
3.31 L2 +x2

3.32 wx > 363 rad/s or 3470 rpm

3.33 (c) 4.462 ft vs 4.570 ft (only 1 inch different) (d) 3.513 ft vs 3.789 ft
(over 3 inches different)

3.34 x = 17.7 ft, the ball is still in the end zone!

3.35 x = 24.0 ft; no, still in the end zone

√
3.36 x = 27 ft, about 5 feet away

3.37 (a) 6 ft in from left post (b) 6 ft in from near post (c) 6 ft in from post
unless kicker more than 6 ft in from post; then, even with kicker

Chapter 4: Collisions
4.1 (a) 220*18/320 = 12.375 mph (b) (220/32)*(18*5280/3600)/.2 = 907.5
lb (c) 907.5*2 = 1815 lb (d) 907.5*3/2 = 1361.25 lb

4.2 (a) defender wins (b) 6/7 mph (c) 707 lb (d) 1414 lb (e) 1060.5 lb

4.3 (a) 200 lb-s (b) 600 lb-s (c) 1800 lb-s (d) 800 lb-s

4.4 tendon (a), ball (a) since x2 > x4 for 0 < x < 1

4.5 (a) .816 to .882; a little slow (b) .773 to .842; very slow (c) .728 to .762;
very slow (d) .693 to .728; very slow

4.6 116 mph; no, the club moves forward

√
4.7 from height h, v = 2hg
 Answers and Selected Solutions 233

4.8 (a) 115.7 mph (b) 114.7 mph (c) bat speed

4.9 (a) 144.7 mph (b) 166.9 mph

4.10 (mbat vbat − mball ((1 − COR)vball + CORvbat ))/(mbat + mball )

4.11 (a) decreases (b) increases if mbat > mball COR

4.12 (a) 55.3 mph (b) 70.9 mph

4.13 To increase impact time, (a) pull hands back (b) bend knees and roll (c)
start catching ball on way up (d) pull racket back

4.14 Greater impact time means sand absorbs more speed

4.15 looser strings give more power (trampoline effect), less control

4.16 decreases dramatically

4.17 off-center hit causes racket to twist; amount of twist affected by grip

4.18 a huge ball is hard to hit

4.19 B exits with more horizontal velocity so the angle is smaller

4.20 Yes if spin velocity high enough; topspin makes balls harder to reach
because of higher horizontal velocity

4.21 lower; topspin; ball more likely to drop into basket

4.22 higher; backspin; volley can go higher/farther than expected

4.23 backspin; increase

4.24 wball up as vball increases; curveball; more backspin, more distance

4.25 club would dig into ground and ball would not go far; club hitting ball
produces backspin which increases distance

4.26 larger COR; less friction; large rating

4.27 fingertips give more, creating more time and less force, making it easier
to catch the ball

4.28 (a) impulse 133.3, force 666.6 (b) impulse 123.7, force 636.2
Rb
4.29 a
(abt − at2 )dt = 23 bab2 /4

4.31 (a) .927 (b) .768 (c) .940 (d) .845

234 Answers and Selected Solutions

4.32 (mball (1 + COR)(vbat + vball )/(mbat + mball )2 > 0 so a larger mbat pro-
duces a larger wball

4.33 −(mbat (1 + COR)(vbat + vball )/(mbat + mball )2 < 0 so a larger mball

produces a smaller wball

4.34 (a) positive, larger bat gives larger bat speed (b) negative, larger ball
gives smaller bat speed

4.35 increase in vbat of 1 increases wball by 1.2

4.36 (a) 238mball /(mball + mbat )2 > 0 so larger bat gives more ball speed (b)
−238mbat /(mball + mbat )2 < 0 so larger ball gives less ball speed

Chapter 5: Ratings Systems

   
5 −2 −1 −2 5 5 −2 −1 −2 24
−2 5 −2 −1 −1 −2 5 −2 −1 5 
5.1 (a) 
−1
 (b)  
−2 5 −2 1 −1 −2 5 −2 5 
−2 −1 −2 5 −5 −2 −1 −2 5 −34
 
7 −2 −1 −2 7/2
−2 7 −2 −1 1/2 
(c) 
−1

−2 7 −2 3/2 
−2 −1 −2 7 −3/2

5.2 Massey win A: 1.42, B: 0.67, C: 0.75, D: 0; Massey points A: 8.46, B: 6.5,
C: 5.29, D: 0; Colley A: .775, B: .475, C: .525, D: .225; Massey points has B
over C. Reasonable c values from 6 to 10. The points and win rankings are
different.
   
6 −2 −1 −3 4 6 −2 −1 −3 16
−2 4 −2 0 0 −2 4 −2 0 0 
5.3 (a) −1 −2 5 −2 1  (b) −1 −2 5 −2
  
6 
−3 0 −2 5 −5 −3 0 −2 5 −22
 
8 −2 −1 −3 3
−2 6 −2 0 1 
(c) 
−1 −2 7 −2 3/2 


−3 0 −2 7 −3/2

5.4 Massey win A: 1.13, B: 0.97, C: 0.81, D: 0; Massey points A: 4.75, B: 4.31,
C: 3.88, D: 0; Colley A: .67, B: .57, C: .54, D: .23; A, B, and C are now rated
nearly the same, although the ranking remains A/B/C.
 Answers and Selected Solutions 235

5.5 5Ao − 2Bd − 3Dd = 72, 4Bo − 2Ad − 2Cd = 44, 4Co − 2Bd − 2Dd = 40,
5Do − 3Ad − 2Cd = 40, −5Ad + 2Bo + 3Do = 50, −4Bd + 2Ao + 2Co = 44,
−4Cd + 2Bo + 2Do = 40, −5Dd + 3Ao + 2Co = 62, Ad + Bd + Cd + Dd = 0.

5.6 A: 13.82, B: 11.82, C: 9.82, D: 7.82; all have 7.82 added to old rating (so
that the average rating is 10.82, the average number of runs scored)

5.7 (a) A 14, B 12 (b) A 15, C 11 (c) A 14, D 8 (d) B 11, C 9 (e) B 11, D 7
(f) C 10, D 8
 
2 −2 0 0 2
−2 2 0 0 −2
5.8 No connection between AB and CD; no; (a)  
0 0 2 −2 0 
  0 0 −2 2 0
4 −2 0 0 2
−2 4 0 0 0
(b)  
0 0 4 −2 1
0 0 −2 4 1

5.9 (a) two rows of zeros, no column of −1s; a − b = 1, c − d = 0; A beats B,

C and D equal (b) a = 3/4, b = 1/4, c = 1/2, d = 1/2; 2 wins boost A from
1/2 to 3/4, drop B to 1/4; C and D stay at 1/2
   
3 −2 −1 0 1 5 −2 −1 0 3/2
−2 2 0 −2
0  (b) −2 4 0 0 0 

5.10 (a) 
−1 0
 (c) C,D
3 −2 1  −1 0 5 −2 3/2
0 0 −2 2 0 0 0 −2 4 1
should rate higher than A,B

5.11 (a) C(0), D(0), A(−1), B(−2) (b) C(.63), D(.57), A(.53), B(.27); Colley
gave C more credit for beating A than Massey

5.12 Scoring: Baylor, TCU, Marshall, Oregon, Ohio State (Marshall had bad
SOS; Georgia was 8th, had good SOS); Defense: Ole Miss, Stanford, LSU,
Alabama, Memphis (Memphis had bad SOS)
 
3 0 0 0 0 −1 −1 −1 34
0
 3 0 0 −1 0 −1 −1 12  
0
 0 3 0 −1 −1 0 − 1 −3 

0 0 0 3 −1 −1 −1 0 4 
5.13 
  A(14.2), B(−3.3),
 0 −1 −1 −1 3 0 0 0 −10 

−1 0 −1 −1 0 3 0 0 3 
 
−1 −1 0 −1 0 0 3 0 −20
−1 −1 −1 0 0 0 0 3 −20
C(0.6), D(4.2); BR better because in general home teams are better than
road teams so BR has better SOS; DR plays AH, best team in league; AR
does not
236 Answers and Selected Solutions

5.14 Column sums are 2, 2, 2, ..., n so 2r1 + 2r2 + ... + 2rn = n or r1 + r2 +

... + rn = n/2 and (r1 + r2 + ... + rn )/n = 1/2.

5.15 change is k(s1 − m1 ) + k(s2 − m2 ) = k(s1 − m1 + (1 − s1 ) − (1 − m1 ) = 0

5.16 (a) 2802.4 and 2597.6 (b) 2792.4 and 2607.6

1 10n 1
5.17 1+10−n = 10n +1 which equals 10n times 1+10n

5.18 2605 and 2595 and then 2599.9 and 2600.1; not the same.

5.19 should stay 7; increases to 7.1; to 7.3; weird

5.20 (a) 125.8 and 84.2 (b) 93.8 and 116.2

5.21 (a) change 15.9 or 16.1 (b) change 15.6 or 16.4

5.24 weight recent games more than early games

5.25 more teams get recognition, leaving a team out is more harmful, voters
need to know more

5.26 (a) rating (b) ranking (c) ranking (d) ranking

5.28 where/when games are played, injuries, new starters

5.29 wins; right

5.30 Elo has equal numerical changes, but higher percentage change for un-
derdog

5.31 opponents’ ratings are part of equations

5.32 no, all numbers in equations double; yes, more games with which to
change Elo ratings
  
38 0 10 0 0 −18 0 −30 634
 0
 24 0 10 −18 0 −16 0 
  356 
 
 10
 0 24 0 0 −16 0 −18   372 
 
 0 10 0 38 −30 0 −18 0   430 
5.34 
   
 0 −18 0 −30 39 1 11  −458
1   
−18 0 −16 0 1 25 1 11  −400
  
 0 −16 0 −18 11 1 25 1  −325
−30 0 −18 0 1 11 1 39 −606

5.35 Sum of rows is zero, so the rows are dependent. If A plays B, then B
plays A; if A wins by 4, then B loses by 4; the sums are zero.

5.37 (a) 0.7345 = 0.213 (b) 0.212

5.38 (a) schedule a (b) schedule b (c) 0.88 versus 0.94, schedule a
 Answers and Selected Solutions 237

Chapter 6: Voting Systems

6.1 (a) Yes, Bradford would still win with 900+315=1215 points, to McCoy’s
798+288=1086 and Tebow’s 927+207=1134, but Tebow would finish second.
(b) With x = 20, Tebow gets 20*309+414+234=6828 points to Bradford’s
6826. (c) In 2008, Bradford would have had 300+315+196=811 votes to Mc-
Coy’s 784 and Tebow’s 750.

6.2 No, if e.g. he receives no 2nd or 3rd place votes.

6.3 (a) Alexander 10 2nd place votes for 48 points, Manning 30 2nd place
votes for 56 points. (b) Alexander 10 2nd place votes for 48 points, Brady 30
2nd palce votes for 50 votes.

6.4 Rodriguez 7, Martinez 4; Martinez was a pitcher.

6.5 Anchorage dropped from 23 to 22 votes. New information.

6.6 (a) Lillehammer (b) Anchorage

6.7 Anchorage beats Lillehammer, 28 to 56.

6.8 (a) A with 8 (b) B with 21 (c) C with 11 (d) no (e) B with 16

6.9 (a) C (b) A (c) A (d) C last place in a majority of lists, but wins plurality.

6.10 (a) C (b) C (c) C (d) B, violates IIA

6.11 (a) 8 A/B/C, 7 B/C/A, 6 C/B/A: A plurality, B Condorcet (b) 10

A/B/C, 5 B/C/A, 4 C/B/A: B Borda, A Condorcet (c) same as (b) with two
approved (d) same as (b) with points 3-2-1

6.12 Higher rated team always wins.

6.13 (a) A wins instead (b) C wins (c) violates IIA

6.14 (a) Number of first place votes (b) Order of elimination (c) Borda yes,
others no

6.15 Votes differ by m (number of voters) for each candidate. 13 A/B/C, 4

B/A/C, 5 B/C/A

6.16 (a) yes (b) yes (c) no (d) no

6.17 In 6.1, change 2 of 11 voters in last column to order B/C/A: B wins.

6.18 Plurality-no, more first place votes helps, Borda-no, more points helps,
approval-no, more votes helps, range-no, more points helps
238 Answers and Selected Solutions

6.19 (a) A with 3 (b) A with 22 (c) 4th highest is 2 (d) B beats C, 5 to 4 (e) A
wins, 7 to 8 (f) A:2,1,1,3,2,3,5; B:3,3,2,1,1,5,3; C:4,2,3,2,3,1,1 (g) medians 2,3;
C beats A, 16 to 17, B is third (h) Third place to first place is unreasonable.

6.20 (a) [1,0,0,0,0,0] (b) [10,7,5,3,1,0] (c) [7,4,3,2,1,0]

6.21 (a) x > 5 (b) x < 5 (c) never

6.22 (a) B gets one of A’s two links and one of D’s three links

6.23 (a) link to A from B 1/2 the time,

 always
 from C, from D 1/3 of the 
time
−1 1/2 1 1/4 0 0 −1 1/2 1 1/4 1/5 0
1/2 −1 0 1/4 0 0 1/2 −1 0 1/4 1/5 0
   
(b) 1/2
 0 −1 1/4 0 0 (c) 1/2
  0 −1 1/4 1/5 0 
 0 1/2 0 −1 0 0  0 1/2 0 −1 1/5 0
0 0 0 1/4 −1 0 0 0 0 1/4 −4/5 0
      
a 0 1/2 1 1/4 1/5 a 1/2
b 1/2 0 0 1/4 1/5 b 1/4
      
(d)  c  = .85 1/2 0 0 1/4 1/5   c  + .15 1/4
     
d  0 1/2 0 0 1/5  d   0 
e 0 0 0 1/4 1/5 e 0

6.24 (a) 0.268 (b) 0.467 (c) 0.070 (d) 0.232 (e) 0.398

6.25 answers vary

6.26 Votes for a given player would be unaffected by votes for or against
others. The negative player is replaced by B on enough ballots, the positive
player replaces A on enough ballots.

6.27 answers vary

6.28 A beats B, B beats C, C beats A; equal but different

6.29 (a) e.g., Condorcet winner or number of 2nd place votes (b) e.g., Con-
dorcet winner or number of 1st place votes

6.30 Yes, but the winner cannot be part of the Condorcet cycle.

6.31 (a) False - might not have majority of 1st place votes (b) True - will
beat everyone (c) False - might not have enough points (d) False - might not
have enough points

6.32 Place B last; only approve of A; 4 stars to A, 1 to others; all are subject
to strategic voting

6.33 The slime mold changes behavior when 3rd alternative presented.
 Answers and Selected Solutions 239

6.34 The existence of medium size shouldn’t affect choice of large or small.

6.35 answers vary

6.36 A majority could prefer A over B, a (different) majority prefer B over

C, and a majority prefer C over A.

6.37 answers vary

6.38 (a) 61 (b) 2 (c) 81

6.39 answers vary

6.40 (a) Voters can’t determine who is best (b) The more voters, the more
the randomness evens out.

6.41 (a) Give strong candidate the worst rating. (b) One biased or extreme
judge in each direction gets ignored. (c) Yes, if too extreme or biased; yes, if
only mimicking other judges.

6.42 One avoids negatives, the other maximizes positives.

6.43 An arbitrary choice that makes the results look good.

6.44 9-seeds win more 1st rounds, but 11-seeds win more 2nd rounds.

6.45 Suppose that A and B have the same number of wins, and that A beat
B. In their other games, then, B won one more game than A, and therefore
beat somebody (call it C) that A did not. The Condorcet cycle is A beat B,
B beat C, and C beat A.

6.46 (a) Let A have the most wins (or tied for the most). For any other team
B, either A beat B or B beat A. If B beat A, then in games against other
teams A won more games than B, so there was at least one team C such that
A beat C and C beat A. (b) If there is a complete Condorcet cycle (A beats
B beats C ... beats A) then everybody wins a medal.

6.47 At least 3.

6.48 Ranking A first puts you left of the 25, so you prefer B to C. Ranking
C first puts you right of the 45, so you prefer B to A. If a majority prefers A
to B, the order A/B/C has a majority of votes, so A beats C. For a majority
to prefer C to B, the order C/B/A must have a majority, so C beats A. If
neither of these occur, B defeats both A and C.

6.49 The only orders are A/B and B/A, one of which has a majority, which
also wins plurality and the Borda count. The majority winner might not be
rated highly enough or approved of by enough voters to win.
240 Answers and Selected Solutions

6.50 Each wins with probability 2/3.

6.51 Yes, if there are 5 votes currently.

6.52 A: 0.228, B: 0.261, C: 0.163, D: 0.196, E: 0.152

6.53 A: 0.196, B: 0.204, C: 0.225, D: 0.189, E: 0.186

6.54 log2 (1) = 0, log2 (2) = 1, log2 (3) = 1.58 ↑ 2, log2 (4) = 2, log2 (5) =
2.32 ↑ 3 and so on. If w is the expected number of wins, w equals 7 − g where
g is the rounded log value.

Chapter 7: Saber- and Other Metrics

7.1 The Cubs’ winning proportion is predicted to be 7012 /(7012 + 7192 ) =
.487. Multiplying by 162 games, the expected number of wins rounds to 79.

7.2 Bad luck, better than their record showed.

7.3 93 > 82; good luck winning close games; no

7.4 exactly 85; very different; good luck didn’t persist

7.5 .553, .573, .474, .409, .390; Los Angeles; yes; Pythagorean better for 3 of
5

7.6 Oakland −11 in 2014: 0, 2, −3, −4, −6, −1, −3, 8, −5 (usually negative);
St. Louis 7 in 2014: −4, −5, 2, −5, 2, −5, 0, 0, 7, 1, 2 (average of 0)

7.7 Cleveland 0.271, Golden State 0.146, large Cleveland advantage

7.8 rapid decrease since 2007

7.9 walks steady decline since 2009; home runs decline since 2000; hits big
decreases 2006-2010, 2014

7.10 slow increase with yearly ups and downs

7.11 rushing up and down, no trend; first downs increase since 2010

7.12 increase since 2004-05

7.13 field goals increase until 2009-10, then decrease; free throw small varia-
tions, no trend
 Answers and Selected Solutions 241

7.14 (a) better offense 28-17, defense 29-16 (b) offense 28-17, defense 24-21
(c) offense 28-17, defense 28-17

7.15 .978; 534 (60 less); Hardy

7.16 .971 is lower; Aparicio made 100 more plays; number of ground balls hit
by Baltimore opponents

7.17 0.44 pt/min and 0.39 pt/min; group scored more with Bogut; quality of
opponents

7.18 (a) 209/150 = 1.39 (b) 0.72 (c) 0.81 (d) 0.94

7.19 (a) 1.32 (b) 1.07 (c) 0.92 (d) 0.98

7.20 Cabrera linear weights 64.4 (Trout 55.1) and wOBA .457 (Trout .425)
the better hitter

7.21 (a) Warriors 0.5 (+.040), 0.123 (+.013), 0.256 (−.004), 0.227 (+.89) won
three, tied one (b) Warriors 0.51 (+.080), 0.091 (−.073), 0.149 (−.184), 0.212
(−.117) won two decisively, lost two decisively

7.22 Ovechkin: 326, 467, 16.1; Kane: 166, 220, 11.0

7.23 (a) Meredith 87.7, Jurgensen 84.5 (b) 1966: 64.2, 2014: 87.1

7.24 53.66, 151.67

7.25 158.33; Romo’s stats in exercise 7.24 with 250 yards instead of 218

7.26 0.78 ∗ 1780 = 1388 close to 1383

7.27 0.831 ∗ 4.2 = 3.49 close to 3.44

7.28 c = 1: x − (x − a) = a; c = 0: x − 0 = x; smaller

7.29 relief pitching, clutch hitting, several blow-out games; continues if clutch
hitting and pitching are repeatable skills

7.30 more playing time for players who walk often; if all teams get walks, no
advantage gained

7.31 since OBP was more predictive of success, it is likely that OBP was
undervalued

7.32 Points for a basketball sub who scores against the other team’s subs.
Points per minute when all players get similar playing times

7.33 a persistent statistic is predictive of the future and indicates a stable

skill
242 Answers and Selected Solutions

7.34 putting has large element of luck, under pressure putting efficiency can
decrease

7.35 it may measure the role of the player or team strategy more than a skill

7.36 winning depends on offense and defense of players not under the pitcher’s
control

7.37 Range Factor gives credit to fast defenders who make many plays, does
not over-penalize errors

7.38 depends on quality of teammates and opposition, and the scoring pace
of the teams

7.39 both give more credit to 3-pointers than 2-pointers, TSP accounts for
scoring by free throws

7.40 each offensive rebound gives the team an extra shot for its one possession

7.41 the other team gets offensive rebounds, some shots go out of bounds
without a rebound

7.42 it limits the impact of an extreme stat in a game

7.43 the relationships among the variables could change with new strategies,
rules, players

7.44 trades and injuries make a team better or worse; these factors would not
be different for even- and odd-numbered games

7.45 praise follows an unusually good play, yelling after an unusually bad
play; the next plays are likely to be more normal. Yelling seems to work.

7.46 (a) a worse score (b) a worse game by regression to the mean

7.47 short passes are automatic completions and may not help the team score

7.48 1.81

7.49 (a) 3.01 (b) 14.51; larger exponents for higher scoring games

7.50 (a) 2.03 for points (b) 1.26 for points; less accurate because of increased
importance of one goal

7.51 Rothman 0.1552 worse than Pythagorean 0.1439

−2x2 y
7.52 wy = (x2 +y 2 )2
 Answers and Selected Solutions 243

7.53 NFL x = y = 361 and a = 3/01 gives 0.5 + 0.00208(x − y); NBA
x = y = 8201 and a = 14.51 gives 0.5 + 0.000442(x − y); NHL x = y = 218
and a = 2.03 gives 0.5 + 0.00233(x − y); coefficients are higher

7.54 (a) 2004 more linear (b) both very scattered (c) 2004 more linear

7.55 (a),(c) both very scattered (d) has tighter grouping than (b)

7.56 (a) has tighter grouping than (b); (c), (d) both very scattered

7.57 (a) 0.22 (b) 0.71 (c) 0.50 (d) 0.76

7.58 (a) 0.020 (b) 0.055; neither correlation is high but possession is higher

7.59 (a) 61.7 (b) 73.1 (c) 74.7 (d) 76.6 (e) 84.1; passing more often and more
efficiently

7.60 −143.33+5.4136shots−0.008124possess+1.1038pass+3.1596fouls’ shots

and passing important, possession not, more fouls go with increased scoring

7.61 4x − 3.3; data is from a parabola y = x2 not a line

Chapter 8: Randomness in Sports

8.1 The means and standard deviations for the American League East, Cen-
tral, West, and National East, Central, and West are, respectively: ALE - 82.2,
9.31; ALC - 81.4, 9.29; ALW - 82.0, 13.10. The largest mean (best division) is
the East. The smallest standard deviation (most balance) is the Central.

8.2 (a) 13 (b) 2 (c) 0

8.3 (a) B (b) B (c) A

8.4 (a) peak 30-35, equal numbers to the left and right but more spread out
to the left (b) biggest peak 40-45, spread out to the left with another peak
20-25; Patriots had a wider spread

8.5 (a) nearly uniform distribution (b) somewhat normal distribution, too
many values 15-20. The NBA seems to be normally distributed around the
50-50 mark, the NFL has more good/bad teams

8.6 won 2 versus spread; mean 0.2, standard deviation 9.3

8.7 84%
244 Answers and Selected Solutions

8.8 (a) 0.154 (b) σ ≈ 5, 16%

8.9 (a) 0.026 (b) σ = 4, 16%

8.10 (a) 0.00012 (b) 0.046 (c) more likely

8.11 (a) 0.00058 (b) 0.0059 (c) very unlikely (1 in 170)

8.12 (a) 91 (b) σ ≈ 6.3, 0.5 (c) 0.16

8.13 (a) 0.5301 and 0.53 (b) 0.5309 and 0.53 (c) 0.5444 and 0.53 (d) 0.6 and
0.6 (e) 0.5990 and 0.6

8.14 (a) Gini 0.44, entropy 1.276 (b) Gini 0.04, entropy 1.606

8.15 (a) Gini 0.37, entropy 2.050 (b) Gini 0.06, entropy 2.298

8.16 max 0.9, min 0

8.17 0.471

8.18 (a) 0.16 (b) 0.025 (c) Smaller σ so fewer large values

8.19 means: 148.3, 148.1, 156.6, 149.9, 146.1, 149.7, 147.6, 144.5; sigmas: 1.99,
1.98, 2.59, 1.86, 1.70, 2.51, 0.80, 1.03; no, but the courses change; 3 of the last
4 have been low; yes, the last two especially; based on the standard deviations,
runners are better in general

8.20 0.027; probably not

8.21 independence: mean 11, variance 4.7; (a) 12 runs, fine (b) 7 runs, bor-
derline (2 sigmas below mean)

8.22 (a) µ = 6, σ = 1, 7 runs is acceptable (b) µ = 9.4, σ = 1.8, 11 runs is

acceptable

8.23 µ = 9, σ = 1.5, 5 runs is too few; 0.002

8.24 µ = 207, σ = 9.7, 206 is acceptable. Streak lengths are close to expec-
tation. No evidence. If p = 159/451, the probability of 6 in a row is 0.0019.
Given enough time, unusual things occur.

8.25 (a) 1 − 0.654 = 0.82 (b) 1 − 0.655 = .88

8.26 (a) 0.9375 (b) 0.937556 = 0.027

8.27 (a) 0.424 (b) 0.331

8.28 (1−.8456 )25 is the probability of failing 25 times to get a 56-game streak.
0.001436 vs 0.001437
 Answers and Selected Solutions 245

8.29 a probability of 2 cannot be correct; 1 − (1 − .001)2000 = 0.865

8.30 Kershaw .252 and 2.56 in 2014, .264 and 3.06 in 2013; Maddux .254 and
2.42 in 1994, .272 and 3.07 in 1993. BABIPs all better than average, Maddux
least lucky in 1993.

8.31 Gwynn .388 and .364, well above average; Thome .353 and .319, above
average

8.32 smaller; the closer p is to 1 the less chance of failures; also, p(1 − p) has
a max value at p = .5

8.33 Improved technology can extend the performance limits.

8.34 balance fair competition with having recognizable stars

8.35 analytics has shown that batting average is not the best stat, so it is not
selected for as rigidly

8.36 Boggs had to contend with better bullpens, more travel, but Williams
had worse conditions

8.37 At possible last games (4-7), the 2-3-2 and 2-2-1-1-1 models have the
same split of home and road games. The 4-3 model starts with a 4-0 split
instead of 2-2; the 2-1-2-2-2 model starts with a 3-1 split.

8.38 Players who break serve may be better players, and more likely to hold
serve for that reason.

8.39 at each step, H and T are equally likely; the runs test tests the pattern
of streaks against expected patterns

8.40 the feeling of being unstoppable and the game becoming easy with suc-
cess automatic; answers vary

8.41 (a) too many H’s (b) no long streaks (c) no streaks of length 1

8.44 could be coincidence, but humans pay attention to round numbers

8.45 batters who hit line drives have higher BABIP, pitchers are affected by
quality of defense (which does not change)

8.46 not informative, since Clemson was also unbeaten in its other uniforms

8.47 (a) peaks 95-100, 100-105; symmetric (b) peaks 90-95, 100-105; bell
shape (c) peak at 2, long tail to the right (d) peaks at 1,2, shorter tail to
the right
246 Answers and Selected Solutions

8.48 (a) decreasing (1/x like) from 92 with 2-3 to 55 with 4-5 to 28 with 12-13
and so on (data from baseball-reference.com) (b) decreasing from 96 with 2
to 78 with 3 to 17 with 7 and so on

8.49 0.6827; this is the 68% of the empirical rule

8.50 (a) 20x−2 (b) 20e−x

8.51 about 3.9 points

8.52 8(b) 0.179 with 55, 0.131 without, averages 16%; 9(b) 0.189 with 76,
0.131 without, averages 16%

8.53 20

12 8 11 7
12 p (1 − p) ; p (1 − p) (12(1 − p) − 8p = 0 so p = 0.6

8.54 1/3

8.55 If f (p) = −plnp − (1 − p)ln(1 − p), f 0 (p) = 0 if p = 0.5.

Xn
−(1/n)ln(1/n) = ln(n)
i=0

8.56 0.530 (2010), 0.523 (2000), 0.567 (1990), 0.492 (1980); no trends appar-
ent

8.57 (a) both the same: 5.876 (b) 2-3-2: 5.786, 2-2-1-1-1: 5.723 (shorter)

8.58 36, since 33 at bats with at least 11 hits for a .200 hitter has probability
0.051.

8.59 (a) 0.09 (b) 0.051 (almost significant)

8.60 (a) 0.739 (b) 0.742

8.61 The probability of getting 4 or less from a binomial variable with pa-
rameters n = 26 and p = .54 is .000056.

Chapter 9: Sports Strategies

9.1 (a) p(80) = 6 − .07(80) = 0.4 (b) p(90) = 6 − .07(90) = −0.3 (c) p(x) = 0
if x = 6/0.07 ≈ 86, at own 14. (d) Each team equally likely to score.

9.2 (a) −1.727 (go for it), −2.22 (punt) (b) −2.427 (go for it), −2.92 (punt)

9.3 p = .42
 Answers and Selected Solutions 247

9.4 p = .57

9.5 p = .28

9.6 2-point conversion, 1.06 > 1

9.7 p(50) − p(80) = −.07(50) − −.07(80) = 2.1 points; p(x − 30) − p(x) =
−.07(x − 30) + .07x = 2.1 points
 
.2 .3 .5 0 0
.2 .3 0 .5 0 
   
9.8 (a)  0 0 .2 .3 .5 (b) .2, .7, .2, .7 (c) a = .2 + .2a + .3b + .5c,

 0 0 .2 .3 .5
0 0 0 0 1
b = .7 + .2a + .3b + .5d, c = .2 + .2c + .3d, d = .7 + .2c + .3d
(d) 1.55, 2.3, 0.7, 1.2

9.9 .573 versus .344

9.10 .637 versus .610

 
0 .6 .4 0 0 0
0 0 0 .3 .7 0 
 
0 0 0 .4 0 .6
9.11 
0

 0 0 0 .6 .4
0 0 0 0 1 0
0 0 0 0 0 1

9.12 (a) 27.1% (b) 88.7%

9.13 The batter does worse against the slider, whether guessing fastball or
slider. Since the pitcher’s slider strategy dominates the fastball strategy, he
should always throw the slider. Knowing this, the batter should always guess
slider and will bat .300.

9.14 The strategies are mixed (random). For the batter, .4p + .25(1 − p) =
.2p + .3(1 − p) gives p = .2: batter guess fastball 20%, slider 80%. For the
pitcher, .4p + .2(1 − p) = .25p + .3(1 − p) gives p = .4: pitcher throws fastball
40%, slider 60%.

9.15 (a) 26 (b) 25 (c) 24-26

9.16 The correlation of best and wins is 0.39, and the correlation of worst
and wins is 0.50.

9.17 The correlation of starters and wins is 0.75 and the correlation of subs
and wins is 0.56.
248 Answers and Selected Solutions

9.18 (a) For events in close games, win probability may be more important.
(b) Mediocre teams play many close games, with many opportunities to gain
win probability points.

9.19 The expected points lost are the same at all locations. With 10 seconds
left, a fumble at the 50 could be less harmful than one near the goal line.

9.20 Flip a coin or use a roulette wheel.

9.21 Tradition and potential ridicule for standing still.

9.22 There is a risk of missing by kicking too high.

9.23 Kelley’s strategy is high variance in the sense that more points are likely
to be scored.

9.24 They get the same shots as a team that lets the shot clock run down.

9.25 Individual players can dominate basketball and baseball more than high-
level soccer.

9.26 The optimal strategy is a random mix of different types of serves, which
is independent of the situation.
   
.1 .2 .7 0 1.143 .286 .943 .486
.1 .2 0 .7 −1
 .143 1.286 .243 1.186
 0 0 .1 .2 (b) (I−S) =  0
9.28 (a) S =    
0 1.143 .286 
0 0 .1 .2 0 0 .143 1.286

9.29 The equations translate as x = R + Sx where x is the 4x1 vector of

unknowns a, b, c, and d. Then Ix − Sx = R and x = (I − S)−1 R.

9.30 (a) .398 (b) .634 (c) .276 (d) .428 (e) .670 (f) .398 (g) drops .398 to .276
to .134 (h) about 50% more (i) .272 (.331 with 0 outs, .072 with 2 outs)

9.31 (1) (a) 0.081 (b) 0.173 (e) 0.192 (f) 0.074 (2) (a) 0.137 (b) 0.243 (e)
0.256 (f) 0.129

9.32 (a) About 49%, less than the actual value of 60% (b) 93%, closer to the
second serve percentage. A change in rules would eliminate the first serve, not
the second serve.

9.33 Tie: 12.7%, Win: 8.6%

9.34 (a) 0.186 (b) third place to fifth place are separated by less than 0.05,
as are sixth and seventh (c) different sprinters have different strengths

9.35 (a) 0.801 (b) 0.856 (c) 0.848 (d) Ten yards is slightly more valuable in
this situation.
 Answers and Selected Solutions 249

Chapter 10: Big Data and Beyond

10.1 Suppose the opponents all have high catch-and-shoot percentages, low
pull-up shooting percentages, and the point guard has a low shooting percent-
age on drives. The defense should stay close to potential shooters, trying to
force dribbles after the pass or allowing the driver to shoot.

10.2 A player who is measurably slowing down as the match continues can
be subbed out, or an opponent who was slowing down would not have to be
marked as closely.

10.3 A player could see that, for example, the opponent always served to the
backhand and slide over a step, or see that the opponent was returning serves
from well behind the baseline and come in to the net.

10.4 The coach could see if a player was tired and slowing down, or injured
and not cutting sharply, or fully recovered from an injury and moving well.

10.5 (a) the approach shot is +0.36 strokes worse than average, the putts
−0.28 strokes better than average; the overall Strokes Gained is still −0.114,
the difference between the actual score of 4 and the expected score. (b) the
drive is −0.234, the approach shot +0.28, and the putts +0.12, for an overall
+0.166 strokes.

10.6 (a) f (1.98) = 31.1 (b) g(1.3) = 30.0 (c) h(3.4) = 17.0 (d) largest effect
of 7 percentage points for running time, then 6.286 for delivery time, then 2.75
for pop time (each equal to 0.1 times the coefficient of x in the appropriate
formula)

10.7 (a) 241/70 = 3.44 s (b) f (1.8) = 36 and g(1.2) = 36.3 so an estimate of
36% is reasonable (c) Pop time and delivery time are better than average, so
the running time needs to be smaller; 3.44 is too high

10.8 (a) 250 (b) 2500 (c) 2500 extra calls is not realistic

10.9 (a) 5.8409/0.1078 = 54 yards; looks reasonable (b) (ln3 +

5.8409)/0.1078 = 64 yards; still reasonable (c) (ln9 + 5.8409)/0.1078 = 75
yards; not likely

10.10 Kickers who stay employed are so accurate that a small number of
misses separates the best from the worst, so bad luck can have a large effect.

10.11 (a) y = −1 (b) slope 1, point (5,0): y = x − 5 (c) slope −1/2, point
(5,3): y = −0.5(x − 5) + 3
250 Answers and Selected Solutions

10.12 The dividing lines y = x − 5 (see 10.11(b)) and y = 5 − x intersect at

(5,0). Keep the portions of these lines to the right of x = 5 and the portion
of the line y = −1 (see 10.11(a)) to the left of (5, 0).

10.13 (a) A is closest: AP=4.1, BP=4.5, CP=4.2 (b) C is closest: AP=5.4,

BP=4.5, CP=4.2
√
10.14
√ The√convex hull is still the triangle ABC, which now has sides 10,
26, and 20. The area is 7. C is farther from A and B, the defense is more
spread out, and the area is larger.
√ √
√ The convex hull is the triangle ABD, which now has sides 13, 2,
10.15
and 17. The area is 2.5.

10.18 (a) 0.580 (b) 0.581

10.19 At 13-4, B would still be larger so B would need to lose two more games
to finish 12-5.

10.20 (a) Playing out of conference means that A (or B) is not the opponent.
Weaker teams could potentially schedule more easy wins and improve the
C-OC mark. (b) The totals become 155-133 for A and 162-127 for B. B’s
advantage has decreased from 7 games to 6.5 games.

10.21 The reliability is increased by the extra data. A loss of charm could
be overcome by these stats being made readily available on smart phones and
video screens at games.

10.23 Receivers have time to adjust on long passes and can outmaneuver
defenders. There is little time for a receiver to adjust on a short pass.

10.25 Neither addresses the quality of the pitcher’s pick-off move, but slow
delivery and pop times greatly increase the probability of a stolen base being
successful.

10.26 Good framing technique has little motion, as if the pitcher is throwing
to the exact intended spot. Blatant movement of the glove after catching the
ball does not work.

10.27 The differences among kickers are small, so the performance differences
could be largely a matter of luck.

10.28 All rating systems have flaws and blind spots, but this obvious bias is
not good for a system used for important decisions.

10.29 It does not appear easier to rebound a 3-point shot as opposed to a

2-point shot, but this is a different question than long shot versus short shot.
 Answers and Selected Solutions 251

10.30 (a) The denominator has derivative (1 + e−5.8409+0.1078d )0 =

0.1078e−5.8409+0.1078d > 0; since the denominator is increasing, the function is
decreasing. The longer the kick, the lower the probability of making the kick.
1
(b) 1+e−5.8409 = 0.997 (c) 0 (d) the limit as d goes to 0 should be 1 (0.997 is
close)
R π/2 R 15 k 1
10.31 −π/2 1 r r dr dθ = 14kπ so k = 14π . A: 0.42, B: 0.58

10.32 (a) The trapezoid has height 2 and parallel sides of length 5 and 3, so
its area is 8. (b) The areas are 3 for ABC and BCD and 5 for ABD and ACD.
Adding up the areas double counts each region so half the sum gives a total
area of 8.

10.33 (a) 0.163 (b) 0.005 (c) 0.186

10.34 (a) 0.108 (b) no, 2-2 has a probability of 0.122

Index
10000 hour rule, 42
absorbing state, 184
acceleration, 1, 2, 13
angular, 26
Achilles tendons, 61
air drag, 6, 12–14
analytics, 121
angular acceleration, 26
angular momentum, 33
conservation of, 33
angular velocity, 26
approval voting, 99
Archimedes’ rule, 59
area between the curves, 160
Arrow’s Impossibility Theorem, 102
Arrow, Kenneth, 102
average velocity, 2
BABIP, 171
backspin, 9, 66
baseball
batting average, 125
delivery time, 207
fielding percentage, 130
framing pitches, 208
Hall of Fame, 97, 100
HOF, 142
linear weights, 140
on base percentage, 125
park factors, 133
pitches, 204
pop time, 207
Range Factor, 130
stolen bases, 140
tag time, 207
WAR, 141
wOBA, 146
basketball
effective field goal percentage,
134
floor spacer, 213
NCAA, 110
points per 48 minutes, 128
rebounding, 211
rebounding rate, 128
statistics, 204
batting average, 125
BCS, 75
Bernoulli trials, 112, 155
binomial coefficient, 156
binomial distribution, 161
binomial model, 155
binomial probability, 156
Borda count, 99
center of mass, 4, 50
coefficient of restitution, 62–64
cognitive illusion, 166
Colley rating system, 75, 81
collision
elastic, 63
inelastic, 63
conditional probability, 165
Condorcet cycle, 98
Condorcet loser, 114
Condorcet winner, 99
connected, 77
conservation
angular momentum, 33
energy, 46
linear momentum, 56, 63
conversion, 211
convex hull, 214
253
254 Index

COR, 62, 63 forces, 1

correlation, 129 Fosbury flop, 49
Corsi rating, 136 Four Factors, 134
Court Pace Rating, 67 framing pitches, 208
Cy Young Award, 96 free throws, 66

delivery time, 207 game control, 188

differential equation, 7, 12 game theory, 189–194
separable, 8 Gini index, 159
dimples, 11 golf
DIPS, 171 slice, 11
distribution golf clubs
binomial, 155 driver, 31
normal, 152 putter, 32
Poisson, 215 golf handicap system, 92
power law, 153 Google, 108

effective mass, 64 hang in the air, 3

elementary row operations, 88 heat map, 217
elimination, 88 Heisman Trophy, 96
Elo rating system, 75, 83 Heron’s formula, 214
empirical rule, 152 high jump, 49
entropy, 163–164 histogram, 153
equivalent voting systems, 115 hockey
expected points formula, 180 Corsi, 136
expected runs matrix, 187 Fenwick, 136
expected value, 180 plus-minus, 135
eye on the ball, 39–41 hockey stick, 68
eyesight, 42 HOF, 95
hot hand, 166–171
factorial, 156 hustle, 211
fairness criteria, 101
Fenwick rating, 136 IIA, 101
field goal accuracy, 209 impulse, 56, 59–61
field goal angle, 43 impulse-momentum equation, 56, 57,
field goal center of field, 44 60
flash-lag effect, 48 independence of irrelevant
floor spacer, 213 alternatives (IIA), 101
foot compression, 61 independent, 155, 164–166, 171, 173
football soccer goals, 215
field goal accuracy, 209 information theory, 163
fumble, 183 instant runoff voting, 99
gambles, 182 integration, 59
going for it, 180
QBR, 206 James, Bill, 122
Quarterback Rating, 136, 140 Pythagorean Method, 122–125
 Index 255

Range Factor, 130 NCAA basketball tournament, 110

Joe DiMaggio hitting streak, 169–171 neural network, 220
Newton’s Laws
kinetic energy, 45 first, 7
knuckleball, 15–17 second, 1, 4, 6, 17, 28, 56
third, 9, 60, 66
Laplace formula, 81, 83 normal distribution, 86, 152
layout position, 32
least squares ratings, 79 occlusion, 42
leverage, 188 offside call, 46
lift force, 12, 16 on base percentage, 125
linear algebra, 87 OneBase, 184
linear momentum, 35, 56 OPS, 126
conservation, 56 optical error, 47
linear regression, 139–141 optical tracking data, 167
linear weights, 140 ordinal ranking, 115
log5 method, 158
Lorenz curve, 159 PageRank, 108, 109
paradox of skill, 162
Magnus force, 8, 9, 12–14 parity, 158
majority, 96 park factor, 132
Markov chains, 184–186 payoff matrix, 190
mass, 1, 64 penalty kick, 189
effective, 64 perimeter weighting system, 31
Massey home court, 92 persistence, 128
Massey offense and defense, 79 pike position, 32
Massey rating system, 75–80, 86 plurality, 96, 99
matrix reduction, 87 plurality with elimination, 99
mean, 152 plus-minus, 131
binomial, 155 point spread, 154
runs, 168 Poisson ditribution, 215
Moment of inertia, 28–29 pop time, 207
calculation, 30 positional voting systems, 102
momentum positioning, 211
angular, 33 potential energy, 45
linear, 35, 56 preference list, 99
monotonicity, 101 Pythagorean Method, 122–125
multiple regression, 140
MVP, 95 Quarterback Rating Formula, 136,
MLB, 96 140
NBA, 97
NFL, 96 range voting, 106
ranking, 74
NBA, 164 rating, 74
luck, 161 rebounding, 211–213
skill, 161 reduced form, 87
256 Index

regression to the mean, 138, 156 tennis, 164

relative velocity, 62 graphics, 205
rotational motion, 26–28 tennis racket, 65
row reduction, 77 terminal velocity, 7
row reduction algorithm, 88 time series, 142
RPI, 218 topspin, 11, 66
runs, 168 torque, 28, 33, 46
mean, 168 transition matrix, 185
standard deviation, 168 transitive, 75, 97
runs test, 168 trimmed mean, 106
tuck position, 32
sabermetrics, 122
sampling bias, 172 value added, 205
scalar, 2 variance, 152
scaling, 82 vector, 2, 12
scatter plot, 125 velocity, 2, 13
seeding of tournaments, 110 angular, 26
shift of gaze error, 47 relative, 62
situational statistics, 165 Voronoi diagram, 211
slap shot, 67
snowboarding, 32 Wald-Wolfowitz runs test, 168
soccer wallaby tendons, 61
Castrol Performance, 206 weight, 64
offside, 46 linear, 87
speed, 204 logistic, 87
softball pitcher, 41 step, 87
speed, 2 win probability, 183, 188
rotational, 26
translational, 26
standard deviation, 152, 162
binomial, 155
NBA, 154
NFL, 154
runs, 168
statistics
counting, 127
rate, 127
strategic voting, 107
strength of schedule, 85
Strokes Gained, 206
Super Bowl, 182
superstitions, 139
sweet spot, 32

tag time, 207